Merge github.com:grpc/grpc into uberpoll

src/core/ext/transport/chttp2/transport/chttp2_transport.c

@@ -884,14 +884,23 @@ static void write_action_begin_locked(grpc_exec_ctx *exec_ctx, void *gt,
   GPR_TIMER_BEGIN("write_action_begin_locked", 0);
   grpc_chttp2_transport *t = gt;
   GPR_ASSERT(t->write_state != GRPC_CHTTP2_WRITE_STATE_IDLE);
-  if (!t->closed && grpc_chttp2_begin_write(exec_ctx, t)) {
-    set_write_state(exec_ctx, t, GRPC_CHTTP2_WRITE_STATE_WRITING,
-                    "begin writing");
-    grpc_closure_sched(exec_ctx, &t->write_action, GRPC_ERROR_NONE);
-  } else {
-    set_write_state(exec_ctx, t, GRPC_CHTTP2_WRITE_STATE_IDLE,
-                    "begin writing nothing");
-    GRPC_CHTTP2_UNREF_TRANSPORT(exec_ctx, t, "writing");
+  switch (t->closed ? GRPC_CHTTP2_NOTHING_TO_WRITE
+                    : grpc_chttp2_begin_write(exec_ctx, t)) {
+    case GRPC_CHTTP2_NOTHING_TO_WRITE:
+      set_write_state(exec_ctx, t, GRPC_CHTTP2_WRITE_STATE_IDLE,
+                      "begin writing nothing");
+      GRPC_CHTTP2_UNREF_TRANSPORT(exec_ctx, t, "writing");
+      break;
+    case GRPC_CHTTP2_PARTIAL_WRITE:
+      set_write_state(exec_ctx, t, GRPC_CHTTP2_WRITE_STATE_WRITING_WITH_MORE,
+                      "begin writing partial");
+      grpc_closure_sched(exec_ctx, &t->write_action, GRPC_ERROR_NONE);
+      break;
+    case GRPC_CHTTP2_FULL_WRITE:
+      set_write_state(exec_ctx, t, GRPC_CHTTP2_WRITE_STATE_WRITING,
+                      "begin writing");
+      grpc_closure_sched(exec_ctx, &t->write_action, GRPC_ERROR_NONE);
+      break;
   }
   GPR_TIMER_END("write_action_begin_locked", 0);
 }
@@ -2130,27 +2139,29 @@ static void end_all_the_calls(grpc_exec_ctx *exec_ctx, grpc_chttp2_transport *t,
 
 static void update_bdp(grpc_exec_ctx *exec_ctx, grpc_chttp2_transport *t,
                        double bdp_dbl) {
-  uint32_t bdp;
-  if (bdp_dbl <= 0) {
-    bdp = 0;
-  } else if (bdp_dbl > UINT32_MAX) {
-    bdp = UINT32_MAX;
+  int32_t bdp;
+  const int32_t kMinBDP = 128;
+  if (bdp_dbl <= kMinBDP) {
+    bdp = kMinBDP;
+  } else if (bdp_dbl > INT32_MAX) {
+    bdp = INT32_MAX;
   } else {
-    bdp = (uint32_t)(bdp_dbl);
+    bdp = (int32_t)(bdp_dbl);
   }
   int64_t delta =
       (int64_t)bdp -
       (int64_t)t->settings[GRPC_LOCAL_SETTINGS]
                           [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE];
-  if (delta == 0 || (bdp != 0 && delta > -1024 && delta < 1024)) {
+  if (delta == 0 || (delta > -bdp / 10 && delta < bdp / 10)) {
     return;
   }
   if (GRPC_TRACER_ON(grpc_bdp_estimator_trace)) {
     gpr_log(GPR_DEBUG, "%s: update initial window size to %d", t->peer_string,
             (int)bdp);
   }
-  push_setting(exec_ctx, t, GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE, bdp);
-  push_setting(exec_ctx, t, GRPC_CHTTP2_SETTINGS_MAX_FRAME_SIZE, bdp);
+  push_setting(exec_ctx, t, GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE,
+               (uint32_t)bdp);
+  push_setting(exec_ctx, t, GRPC_CHTTP2_SETTINGS_MAX_FRAME_SIZE, (uint32_t)bdp);
 }
 
 static grpc_error *try_http_parsing(grpc_exec_ctx *exec_ctx,
@@ -2543,7 +2554,7 @@ static void incoming_byte_stream_update_flow_control(grpc_exec_ctx *exec_ctx,
                                    add_max_recv_bytes);
     if ((int64_t)s->incoming_window_delta + (int64_t)initial_window_size -
             (int64_t)s->announce_window >
-        (int64_t)initial_window_size / 2) {
+        2 * (int64_t)initial_window_size) {
       write_type = GRPC_CHTTP2_STREAM_WRITE_PIGGYBACK;
     }
     grpc_chttp2_become_writable(exec_ctx, t, s, write_type,

src/core/ext/transport/chttp2/transport/internal.h

@@ -552,9 +552,14 @@ void grpc_chttp2_initiate_write(grpc_exec_ctx *exec_ctx,
                                 grpc_chttp2_transport *t,
                                 bool covered_by_poller, const char *reason);
 
-/** Someone is unlocking the transport mutex: check to see if writes
-    are required, and frame them if so */
-bool grpc_chttp2_begin_write(grpc_exec_ctx *exec_ctx, grpc_chttp2_transport *t);
+typedef enum {
+  GRPC_CHTTP2_NOTHING_TO_WRITE,
+  GRPC_CHTTP2_PARTIAL_WRITE,
+  GRPC_CHTTP2_FULL_WRITE,
+} grpc_chttp2_begin_write_result;
+
+grpc_chttp2_begin_write_result grpc_chttp2_begin_write(
+    grpc_exec_ctx *exec_ctx, grpc_chttp2_transport *t);
 void grpc_chttp2_end_write(grpc_exec_ctx *exec_ctx, grpc_chttp2_transport *t,
                            grpc_error *error);
 

src/core/ext/transport/chttp2/transport/parsing.c

@@ -418,12 +418,7 @@ static grpc_error *update_incoming_window(grpc_exec_ctx *exec_ctx,
 
     GRPC_CHTTP2_FLOW_DEBIT_STREAM_INCOMING_WINDOW_DELTA("parse", t, s,
                                                         incoming_frame_size);
-    if ((int64_t)t->settings[GRPC_SENT_SETTINGS]
-                            [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE] +
-            (int64_t)s->incoming_window_delta - (int64_t)s->announce_window <=
-        (int64_t)t->settings[GRPC_SENT_SETTINGS]
-                            [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE] /
-            2) {
+    if ((int64_t)s->incoming_window_delta - (int64_t)s->announce_window <= 0) {
       grpc_chttp2_become_writable(exec_ctx, t, s,
                                   GRPC_CHTTP2_STREAM_WRITE_INITIATE_UNCOVERED,
                                   "window-update-required");

src/core/ext/transport/chttp2/transport/writing.c

@@ -163,19 +163,22 @@ static bool stream_ref_if_not_destroyed(gpr_refcount *r) {
   return true;
 }
 
+/* How many bytes of incoming flow control would we like to advertise */
 uint32_t grpc_chttp2_target_incoming_window(grpc_chttp2_transport *t) {
-  return (uint32_t)GPR_MAX(
+  return (uint32_t)GPR_MIN(
       (int64_t)((1u << 31) - 1),
       t->stream_total_over_incoming_window +
-          (int64_t)GPR_MAX(
-              t->settings[GRPC_SENT_SETTINGS]
-                         [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE] -
-                  t->stream_total_under_incoming_window,
-              0));
+          t->settings[GRPC_SENT_SETTINGS]
+                     [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE]);
 }
 
-bool grpc_chttp2_begin_write(grpc_exec_ctx *exec_ctx,
-                             grpc_chttp2_transport *t) {
+/* How many bytes would we like to put on the wire during a single syscall */
+static uint32_t target_write_size(grpc_chttp2_transport *t) {
+  return 1024 * 1024;
+}
+
+grpc_chttp2_begin_write_result grpc_chttp2_begin_write(
+    grpc_exec_ctx *exec_ctx, grpc_chttp2_transport *t) {
   grpc_chttp2_stream *s;
 
   GPR_TIMER_BEGIN("grpc_chttp2_begin_write", 0);
@@ -209,9 +212,20 @@ bool grpc_chttp2_begin_write(grpc_exec_ctx *exec_ctx,
     }
   }
 
+  bool partial_write = false;
+
   /* for each grpc_chttp2_stream that's become writable, frame it's data
      (according to available window sizes) and add to the output buffer */
-  while (grpc_chttp2_list_pop_writable_stream(t, &s)) {
+  while (true) {
+    if (t->outbuf.length > target_write_size(t)) {
+      partial_write = true;
+      break;
+    }
+
+    if (!grpc_chttp2_list_pop_writable_stream(t, &s)) {
+      break;
+    }
+
     bool sent_initial_metadata = s->sent_initial_metadata;
     bool now_writing = false;
 
@@ -398,7 +412,9 @@ bool grpc_chttp2_begin_write(grpc_exec_ctx *exec_ctx,
 
   GPR_TIMER_END("grpc_chttp2_begin_write", 0);
 
-  return t->outbuf.count > 0;
+  return t->outbuf.count > 0 ? (partial_write ? GRPC_CHTTP2_PARTIAL_WRITE
+                                              : GRPC_CHTTP2_FULL_WRITE)
+                             : GRPC_CHTTP2_NOTHING_TO_WRITE;
 }
 
 void grpc_chttp2_end_write(grpc_exec_ctx *exec_ctx, grpc_chttp2_transport *t,

src/core/lib/transport/bdp_estimator.c

@@ -44,6 +44,7 @@ void grpc_bdp_estimator_init(grpc_bdp_estimator *estimator, const char *name) {
   estimator->estimate = 65536;
   estimator->ping_state = GRPC_BDP_PING_UNSCHEDULED;
   estimator->name = name;
+  estimator->bw_est = 0;
 }
 
 bool grpc_bdp_estimator_get_estimate(grpc_bdp_estimator *estimator,
@@ -84,16 +85,26 @@ void grpc_bdp_estimator_start_ping(grpc_bdp_estimator *estimator) {
   GPR_ASSERT(estimator->ping_state == GRPC_BDP_PING_SCHEDULED);
   estimator->ping_state = GRPC_BDP_PING_STARTED;
   estimator->accumulator = 0;
+  estimator->ping_start_time = gpr_now(GPR_CLOCK_MONOTONIC);
 }
 
 void grpc_bdp_estimator_complete_ping(grpc_bdp_estimator *estimator) {
+  gpr_timespec dt_ts =
+      gpr_time_sub(gpr_now(GPR_CLOCK_MONOTONIC), estimator->ping_start_time);
+  double dt = (double)dt_ts.tv_sec + 1e-9 * (double)dt_ts.tv_nsec;
+  double bw = dt > 0 ? ((double)estimator->accumulator / dt) : 0;
   if (GRPC_TRACER_ON(grpc_bdp_estimator_trace)) {
-    gpr_log(GPR_DEBUG, "bdp[%s]:complete acc=%" PRId64 " est=%" PRId64,
-            estimator->name, estimator->accumulator, estimator->estimate);
+    gpr_log(GPR_DEBUG, "bdp[%s]:complete acc=%" PRId64 " est=%" PRId64
+                       " dt=%lf bw=%lfMbs bw_est=%lfMbs",
+            estimator->name, estimator->accumulator, estimator->estimate, dt,
+            bw / 125000.0, estimator->bw_est / 125000.0);
   }
   GPR_ASSERT(estimator->ping_state == GRPC_BDP_PING_STARTED);
-  if (estimator->accumulator > 2 * estimator->estimate / 3) {
-    estimator->estimate *= 2;
+  if (estimator->accumulator > 2 * estimator->estimate / 3 &&
+      bw > estimator->bw_est) {
+    estimator->estimate =
+        GPR_MAX(estimator->accumulator, estimator->estimate * 2);
+    estimator->bw_est = bw;
     if (GRPC_TRACER_ON(grpc_bdp_estimator_trace)) {
       gpr_log(GPR_DEBUG, "bdp[%s]: estimate increased to %" PRId64,
               estimator->name, estimator->estimate);

src/core/lib/transport/bdp_estimator.h

@@ -34,6 +34,7 @@
 #ifndef GRPC_CORE_LIB_TRANSPORT_BDP_ESTIMATOR_H
 #define GRPC_CORE_LIB_TRANSPORT_BDP_ESTIMATOR_H
 
+#include <grpc/support/time.h>
 #include <stdbool.h>
 #include <stdint.h>
 #include "src/core/lib/debug/trace.h"
@@ -53,6 +54,8 @@ typedef struct grpc_bdp_estimator {
   grpc_bdp_estimator_ping_state ping_state;
   int64_t accumulator;
   int64_t estimate;
+  gpr_timespec ping_start_time;
+  double bw_est;
   const char *name;
 } grpc_bdp_estimator;
 

src/proto/grpc/testing/control.proto

@@ -52,6 +52,9 @@ enum ServerType {
 enum RpcType {
   UNARY = 0;
   STREAMING = 1;
+  STREAMING_FROM_CLIENT = 2;
+  STREAMING_FROM_SERVER = 3;
+  STREAMING_BOTH_WAYS = 4;
 }
 
 // Parameters of poisson process distribution, which is a good representation

src/proto/grpc/testing/services.proto

@@ -42,9 +42,22 @@ service BenchmarkService {
   // The server returns the client payload as-is.
   rpc UnaryCall(SimpleRequest) returns (SimpleResponse);
 
-  // One request followed by one response.
-  // The server returns the client payload as-is.
+  // Repeated sequence of one request followed by one response.
+  // Should be called streaming ping-pong
+  // The server returns the client payload as-is on each response
   rpc StreamingCall(stream SimpleRequest) returns (stream SimpleResponse);
+
+  // Single-sided unbounded streaming from client to server
+  // The server returns the client payload as-is once the client does WritesDone
+  rpc StreamingFromClient(stream SimpleRequest) returns (SimpleResponse);
+
+  // Single-sided unbounded streaming from server to client
+  // The server repeatedly returns the client payload as-is
+  rpc StreamingFromServer(SimpleRequest) returns (stream SimpleResponse);
+
+  // Two-sided unbounded streaming between server to client
+  // Both sides send the content of their own choice to the other
+  rpc StreamingBothWays(stream SimpleRequest) returns (stream SimpleResponse);
 }
 
 service WorkerService {

src/python/grpcio/grpc/__init__.py

@@ -61,13 +61,12 @@ class Future(six.with_metaclass(abc.ABCMeta)):
     This method does not block.
 
     Returns:
-      True if the computation has not yet begun, will not be allowed to take
-        place, and determination of both was possible without blocking. False
-        under all other circumstances including but not limited to the
-        computation's already having begun, the computation's already having
-        finished, and the computation's having been scheduled for execution on a
-        remote system for which a determination of whether or not it commenced
-        before being cancelled cannot be made without blocking.
+        bool:
+        Returns True if the computation was canceled.
+        Returns False under all other circumstances, for example:
+        1. computation has begun and could not be canceled.
+        2. computation has finished
+        3. computation is scheduled for execution and it is impossible to determine its state without blocking.
     """
         raise NotImplementedError()
 
@@ -78,10 +77,12 @@ class Future(six.with_metaclass(abc.ABCMeta)):
     This method does not block.
 
     Returns:
-      True if the computation was cancelled any time before its result became
-        immediately available. False under all other circumstances including but
-        not limited to this object's cancel method not having been called and
-        the computation's result having become immediately available.
+        bool:
+        Returns True if the computation was cancelled before its result became
+        available.
+        False under all other circumstances, for example:
+        1. computation was not cancelled.
+        2. computation's result is available.
     """
         raise NotImplementedError()
 
@@ -92,9 +93,10 @@ class Future(six.with_metaclass(abc.ABCMeta)):
     This method does not block.
 
     Returns:
-      True if the computation is scheduled to take place in the future or is
-        taking place now, or False if the computation took place in the past or
-        was cancelled.
+        bool:
+        Returns True if the computation is scheduled for execution or currently
+        executing.
+        Returns False if the computation already executed or was cancelled.
     """
         raise NotImplementedError()
 
@@ -105,22 +107,24 @@ class Future(six.with_metaclass(abc.ABCMeta)):
     This method does not block.
 
     Returns:
-      True if the computation is known to have either completed or have been
-        unscheduled or interrupted. False if the computation may possibly be
-        executing or scheduled to execute later.
+        bool:
+        Returns True if the computation already executed or was cancelled.
+        Returns False if the computation is scheduled for execution or currently
+        executing.
+        This is exactly opposite of the running() method's result.
     """
         raise NotImplementedError()
 
     @abc.abstractmethod
     def result(self, timeout=None):
-        """Accesses the outcome of the computation or raises its exception.
+        """Returns the result of the computation or raises its exception.
 
     This method may return immediately or may block.
 
     Args:
       timeout: The length of time in seconds to wait for the computation to
-        finish or be cancelled, or None if this method should block until the
-        computation has finished or is cancelled no matter how long that takes.
+        finish or be cancelled. If None, the call will block until the computations's
+        termination.
 
     Returns:
       The return value of the computation.
@@ -142,12 +146,11 @@ class Future(six.with_metaclass(abc.ABCMeta)):
 
     Args:
       timeout: The length of time in seconds to wait for the computation to
-        terminate or be cancelled, or None if this method should block until
-        the computation is terminated or is cancelled no matter how long that
-        takes.
+        terminate or be cancelled. If None, the call will block until the computations's
+        termination.
 
     Returns:
-      The exception raised by the computation, or None if the computation did
+        The exception raised by the computation, or None if the computation did
         not raise an exception.
 
     Raises:
@@ -165,12 +168,11 @@ class Future(six.with_metaclass(abc.ABCMeta)):
 
     Args:
       timeout: The length of time in seconds to wait for the computation to
-        terminate or be cancelled, or None if this method should block until
-        the computation is terminated or is cancelled no matter how long that
-        takes.
+        terminate or be cancelled. If None, the call will block until the
+        computations's termination.
 
     Returns:
-      The traceback of the exception raised by the computation, or None if the
+        The traceback of the exception raised by the computation, or None if the
         computation did not raise an exception.
 
     Raises:
@@ -260,7 +262,12 @@ class RpcContext(six.with_metaclass(abc.ABCMeta)):
 
     @abc.abstractmethod
     def is_active(self):
-        """Describes whether the RPC is active or has terminated."""
+        """Describes whether the RPC is active or has terminated.
+
+    Returns:
+      bool:
+      True if RPC is active, False otherwise.
+    """
         raise NotImplementedError()
 
     @abc.abstractmethod
@@ -290,8 +297,9 @@ class RpcContext(six.with_metaclass(abc.ABCMeta)):
       callback: A no-parameter callable to be called on RPC termination.
 
     Returns:
-      True if the callback was added and will be called later; False if the
-        callback was not added and will not later be called (because the RPC
+      bool:
+        True if the callback was added and will be called later; False if the
+        callback was not added and will not be called (because the RPC
         already terminated or some other reason).
     """
         raise NotImplementedError()
@@ -305,7 +313,7 @@ class Call(six.with_metaclass(abc.ABCMeta, RpcContext)):
 
     @abc.abstractmethod
     def initial_metadata(self):
-        """Accesses the initial metadata from the service-side of the RPC.
+        """Accesses the initial metadata sent by the server.
 
     This method blocks until the value is available.
 
@@ -316,7 +324,7 @@ class Call(six.with_metaclass(abc.ABCMeta, RpcContext)):
 
     @abc.abstractmethod
     def trailing_metadata(self):
-        """Accesses the trailing metadata from the service-side of the RPC.
+        """Accesses the trailing metadata sent by the server.
 
     This method blocks until the value is available.
 
@@ -327,7 +335,7 @@ class Call(six.with_metaclass(abc.ABCMeta, RpcContext)):
 
     @abc.abstractmethod
     def code(self):
-        """Accesses the status code emitted by the service-side of the RPC.
+        """Accesses the status code sent by the server.
 
     This method blocks until the value is available.
 
@@ -338,7 +346,7 @@ class Call(six.with_metaclass(abc.ABCMeta, RpcContext)):
 
     @abc.abstractmethod
     def details(self):
-        """Accesses the details value emitted by the service-side of the RPC.
+        """Accesses the details sent by the server.
 
     This method blocks until the value is available.
 
@@ -352,10 +360,12 @@ class Call(six.with_metaclass(abc.ABCMeta, RpcContext)):
 
 
 class ChannelCredentials(object):
-    """A value encapsulating the data required to create a secure Channel.
+    """An encapsulation of the data required to create a secure Channel.
 
   This class has no supported interface - it exists to define the type of its
-  instances and its instances exist to be passed to other functions.
+  instances and its instances exist to be passed to other functions. For example,
+  ssl_channel_credentials returns an instance, and secure_channel consumes an
+  instance of this class.
   """
 
     def __init__(self, credentials):
@@ -363,7 +373,7 @@ class ChannelCredentials(object):
 
 
 class CallCredentials(object):
-    """A value encapsulating data asserting an identity over a channel.
+    """An encapsulation of the data required to assert an identity over a channel.
 
   A CallCredentials may be composed with ChannelCredentials to always assert
   identity for every call over that Channel.
@@ -416,7 +426,7 @@ class AuthMetadataPlugin(six.with_metaclass(abc.ABCMeta)):
 
 
 class ServerCredentials(object):
-    """A value encapsulating the data required to open a secure port on a Server.
+    """An encapsulation of the data required to open a secure port on a Server.
 
   This class has no supported interface - it exists to define the type of its
   instances and its instances exist to be passed to other functions.
@@ -430,7 +440,7 @@ class ServerCredentials(object):
 
 
 class UnaryUnaryMultiCallable(six.with_metaclass(abc.ABCMeta)):
-    """Affords invoking a unary-unary RPC."""
+    """Affords invoking a unary-unary RPC from client-side."""
 
     @abc.abstractmethod
     def __call__(self, request, timeout=None, metadata=None, credentials=None):
@@ -486,7 +496,7 @@ class UnaryUnaryMultiCallable(six.with_metaclass(abc.ABCMeta)):
       credentials: An optional CallCredentials for the RPC.
 
     Returns:
-      An object that is both a Call for the RPC and a Future. In the event of
+        An object that is both a Call for the RPC and a Future. In the event of
         RPC completion, the return Call-Future's result value will be the
         response message of the RPC. Should the event terminate with non-OK
         status, the returned Call-Future's exception value will be an RpcError.
@@ -495,7 +505,7 @@ class UnaryUnaryMultiCallable(six.with_metaclass(abc.ABCMeta)):
 
 
 class UnaryStreamMultiCallable(six.with_metaclass(abc.ABCMeta)):
-    """Affords invoking a unary-stream RPC."""
+    """Affords invoking a unary-stream RPC from client-side."""
 
     @abc.abstractmethod
     def __call__(self, request, timeout=None, metadata=None, credentials=None):
@@ -504,12 +514,13 @@ class UnaryStreamMultiCallable(six.with_metaclass(abc.ABCMeta)):
     Args:
       request: The request value for the RPC.
       timeout: An optional duration of time in seconds to allow for the RPC.
+               If None, the timeout is considered infinite.
       metadata: An optional :term:`metadata` to be transmitted to the
         service-side of the RPC.
       credentials: An optional CallCredentials for the RPC.
 
     Returns:
-      An object that is both a Call for the RPC and an iterator of response
+        An object that is both a Call for the RPC and an iterator of response
         values. Drawing response values from the returned Call-iterator may
         raise RpcError indicating termination of the RPC with non-OK status.
     """
@@ -517,7 +528,7 @@ class UnaryStreamMultiCallable(six.with_metaclass(abc.ABCMeta)):
 
 
 class StreamUnaryMultiCallable(six.with_metaclass(abc.ABCMeta)):
-    """Affords invoking a stream-unary RPC in any call style."""
+    """Affords invoking a stream-unary RPC from client-side."""
 
     @abc.abstractmethod
     def __call__(self,
@@ -530,6 +541,7 @@ class StreamUnaryMultiCallable(six.with_metaclass(abc.ABCMeta)):
     Args:
       request_iterator: An iterator that yields request values for the RPC.
       timeout: An optional duration of time in seconds to allow for the RPC.
+               If None, the timeout is considered infinite.
       metadata: Optional :term:`metadata` to be transmitted to the
         service-side of the RPC.
       credentials: An optional CallCredentials for the RPC.
@@ -539,8 +551,8 @@ class StreamUnaryMultiCallable(six.with_metaclass(abc.ABCMeta)):
 
     Raises:
       RpcError: Indicating that the RPC terminated with non-OK status. The
-        raised RpcError will also be a Call for the RPC affording the RPC's
-        metadata, status code, and details.
+        raised RpcError will also implement grpc.Call, affording methods
+        such as metadata, code, and details.
     """
         raise NotImplementedError()
 
@@ -550,17 +562,18 @@ class StreamUnaryMultiCallable(six.with_metaclass(abc.ABCMeta)):
                   timeout=None,
                   metadata=None,
                   credentials=None):
-        """Synchronously invokes the underlying RPC.
+        """Synchronously invokes the underlying RPC on the client.
 
     Args:
       request_iterator: An iterator that yields request values for the RPC.
       timeout: An optional duration of time in seconds to allow for the RPC.
+               If None, the timeout is considered infinite.
       metadata: Optional :term:`metadata` to be transmitted to the
         service-side of the RPC.
       credentials: An optional CallCredentials for the RPC.
 
     Returns:
-      The response value for the RPC and a Call for the RPC.
+      The response value for the RPC and a Call object for the RPC.
 
     Raises:
       RpcError: Indicating that the RPC terminated with non-OK status. The
@@ -575,17 +588,18 @@ class StreamUnaryMultiCallable(six.with_metaclass(abc.ABCMeta)):
                timeout=None,
                metadata=None,
                credentials=None):
-        """Asynchronously invokes the underlying RPC.
+        """Asynchronously invokes the underlying RPC on the client.
 
     Args:
       request_iterator: An iterator that yields request values for the RPC.
       timeout: An optional duration of time in seconds to allow for the RPC.
+               If None, the timeout is considered infinite.
       metadata: Optional :term:`metadata` to be transmitted to the
         service-side of the RPC.
       credentials: An optional CallCredentials for the RPC.
 
     Returns:
-      An object that is both a Call for the RPC and a Future. In the event of
+        An object that is both a Call for the RPC and a Future. In the event of
         RPC completion, the return Call-Future's result value will be the
         response message of the RPC. Should the event terminate with non-OK
         status, the returned Call-Future's exception value will be an RpcError.
@@ -594,7 +608,7 @@ class StreamUnaryMultiCallable(six.with_metaclass(abc.ABCMeta)):
 
 
 class StreamStreamMultiCallable(six.with_metaclass(abc.ABCMeta)):
-    """Affords invoking a stream-stream RPC in any call style."""
+    """Affords invoking a stream-stream RPC on client-side."""
 
     @abc.abstractmethod
     def __call__(self,
@@ -602,17 +616,18 @@ class StreamStreamMultiCallable(six.with_metaclass(abc.ABCMeta)):
                  timeout=None,
                  metadata=None,
                  credentials=None):
-        """Invokes the underlying RPC.
+        """Invokes the underlying RPC on the client.
 
     Args:
       request_iterator: An iterator that yields request values for the RPC.
       timeout: An optional duration of time in seconds to allow for the RPC.
+               if not specified the timeout is considered infinite.
       metadata: Optional :term:`metadata` to be transmitted to the
         service-side of the RPC.
       credentials: An optional CallCredentials for the RPC.
 
     Returns:
-      An object that is both a Call for the RPC and an iterator of response
+        An object that is both a Call for the RPC and an iterator of response
         values. Drawing response values from the returned Call-iterator may
         raise RpcError indicating termination of the RPC with non-OK status.
     """
@@ -623,27 +638,32 @@ class StreamStreamMultiCallable(six.with_metaclass(abc.ABCMeta)):
 
 
 class Channel(six.with_metaclass(abc.ABCMeta)):
-    """Affords RPC invocation via generic methods."""
+    """Affords RPC invocation via generic methods on client-side."""
 
     @abc.abstractmethod
     def subscribe(self, callback, try_to_connect=False):
-        """Subscribes to this Channel's connectivity.
+        """Subscribe to this Channel's connectivity state machine.
+
+    A Channel may be in any of the states described by ChannelConnectivity.
+    This method allows application to monitor the state transitions.
+    The typical use case is to debug or gain better visibility into gRPC
+    runtime's state.
 
     Args:
-      callback: A callable to be invoked and passed a ChannelConnectivity value
-        describing this Channel's connectivity. The callable will be invoked
-        immediately upon subscription and again for every change to this
-        Channel's connectivity thereafter until it is unsubscribed or this
+      callback: A callable to be invoked with ChannelConnectivity argument.
+        ChannelConnectivity describes current state of the channel.
+        The callable will be invoked immediately upon subscription and again for
+        every change to ChannelConnectivity until it is unsubscribed or this
         Channel object goes out of scope.
       try_to_connect: A boolean indicating whether or not this Channel should
-        attempt to connect if it is not already connected and ready to conduct
-        RPCs.
+        attempt to connect immediately. If set to False, gRPC runtime decides
+        when to connect.
     """
         raise NotImplementedError()
 
     @abc.abstractmethod
     def unsubscribe(self, callback):
-        """Unsubscribes a callback from this Channel's connectivity.
+        """Unsubscribes a subscribed callback from this Channel's connectivity.
 
     Args:
       callback: A callable previously registered with this Channel from having
@@ -736,7 +756,7 @@ class ServicerContext(six.with_metaclass(abc.ABCMeta, RpcContext)):
 
     @abc.abstractmethod
     def invocation_metadata(self):
-        """Accesses the metadata from the invocation-side of the RPC.
+        """Accesses the metadata from the sent by the client.
 
     Returns:
       The invocation :term:`metadata`.
@@ -749,15 +769,16 @@ class ServicerContext(six.with_metaclass(abc.ABCMeta, RpcContext)):
 
     Returns:
       A string identifying the peer that invoked the RPC being serviced.
+      The string format is determined by gRPC runtime.
     """
         raise NotImplementedError()
 
     @abc.abstractmethod
     def send_initial_metadata(self, initial_metadata):
-        """Sends the initial metadata value to the invocation-side of the RPC.
+        """Sends the initial metadata value to the client.
 
-    This method need not be called by method implementations if they have no
-    service-side initial metadata to transmit.
+    This method need not be called by implementations if they have no
+    metadata to add to what the gRPC runtime will transmit.
 
     Args:
       initial_metadata: The initial :term:`metadata`.
@@ -766,10 +787,10 @@ class ServicerContext(six.with_metaclass(abc.ABCMeta, RpcContext)):
 
     @abc.abstractmethod
     def set_trailing_metadata(self, trailing_metadata):
-        """Accepts the trailing metadata value of the RPC.
+        """Sends the trailing metadata for the RPC.
 
-    This method need not be called by method implementations if they have no
-    service-side trailing metadata to transmit.
+    This method need not be called by implementations if they have no
+    metadata to add to what the gRPC runtime will transmit.
 
     Args:
       trailing_metadata: The trailing :term:`metadata`.
@@ -778,27 +799,25 @@ class ServicerContext(six.with_metaclass(abc.ABCMeta, RpcContext)):
 
     @abc.abstractmethod
     def set_code(self, code):
-        """Accepts the status code of the RPC.
+        """Sets the value to be used as status code upon RPC completion.
 
     This method need not be called by method implementations if they wish the
     gRPC runtime to determine the status code of the RPC.
 
     Args:
-      code: A StatusCode value to be transmitted to the invocation side of the
-        RPC as the status code of the RPC.
+      code: A StatusCode object to be sent to the client.
     """
         raise NotImplementedError()
 
     @abc.abstractmethod
     def set_details(self, details):
-        """Accepts the service-side details of the RPC.
+        """Sets the value to be used as detail string upon RPC completion.
 
     This method need not be called by method implementations if they have no
     details to transmit.
 
     Args:
-      details: A string to be transmitted to the invocation side of the RPC as
-        the status details of the RPC.
+      details: An arbitrary string to be sent to the client upon completion.
     """
         raise NotImplementedError()
 
@@ -845,7 +864,7 @@ class HandlerCallDetails(six.with_metaclass(abc.ABCMeta)):
     """Describes an RPC that has just arrived for service.
   Attributes:
     method: The method name of the RPC.
-    invocation_metadata: The :term:`metadata` from the invocation side of the RPC.
+    invocation_metadata: The :term:`metadata` sent by the client.
   """
 
 
@@ -854,14 +873,14 @@ class GenericRpcHandler(six.with_metaclass(abc.ABCMeta)):
 
     @abc.abstractmethod
     def service(self, handler_call_details):
-        """Services an RPC (or not).
+        """Returns the handler for servicing the RPC.
 
     Args:
       handler_call_details: A HandlerCallDetails describing the RPC.
 
     Returns:
-      An RpcMethodHandler with which the RPC may be serviced, or None to
-        indicate that this object will not be servicing the RPC.
+      An RpcMethodHandler with which the RPC may be serviced if the implementation
+      chooses to service this RPC, or None otherwise.
     """
         raise NotImplementedError()
 
@@ -870,15 +889,15 @@ class ServiceRpcHandler(six.with_metaclass(abc.ABCMeta, GenericRpcHandler)):
     """An implementation of RPC methods belonging to a service.
 
   A service handles RPC methods with structured names of the form
-  '/Service.Name/Service.MethodX', where 'Service.Name' is the value
-  returned by service_name(), and 'Service.MethodX' is the service method
-  name.  A service can have multiple service methods names, but only a single
+  '/Service.Name/Service.Method', where 'Service.Name' is the value
+  returned by service_name(), and 'Service.Method' is the method
+  name.  A service can have multiple method names, but only a single
   service name.
   """
 
     @abc.abstractmethod
     def service_name(self):
-        """Returns this services name.
+        """Returns this service's name.
 
     Returns:
       The service name.
@@ -900,88 +919,78 @@ class Server(six.with_metaclass(abc.ABCMeta)):
 
     Args:
       generic_rpc_handlers: An iterable of GenericRpcHandlers that will be used
-        to service RPCs after this Server is started.
+      to service RPCs.
     """
         raise NotImplementedError()
 
     @abc.abstractmethod
     def add_insecure_port(self, address):
-        """Reserves a port for insecure RPC service once this Server becomes active.
+        """Opens an insecure port for accepting RPCs.
 
-    This method may only be called before calling this Server's start method is
-    called.
+    This method may only be called before starting the server.
 
     Args:
       address: The address for which to open a port.
+      if the port is 0, or not specified in the address, then gRPC runtime
+      will choose a port.
 
     Returns:
-      An integer port on which RPCs will be serviced after this link has been
-        started. This is typically the same number as the port number contained
-        in the passed address, but will likely be different if the port number
-        contained in the passed address was zero.
+      integer:
+      An integer port on which server will accept RPC requests.
     """
         raise NotImplementedError()
 
     @abc.abstractmethod
     def add_secure_port(self, address, server_credentials):
-        """Reserves a port for secure RPC service after this Server becomes active.
+        """Opens a secure port for accepting RPCs.
 
-    This method may only be called before calling this Server's start method is
-    called.
+    This method may only be called before starting the server.
 
     Args:
       address: The address for which to open a port.
-      server_credentials: A ServerCredentials.
+        if the port is 0, or not specified in the address, then gRPC runtime
+        will choose a port.
+      server_credentials: A ServerCredentials object.
 
     Returns:
-      An integer port on which RPCs will be serviced after this link has been
-        started. This is typically the same number as the port number contained
-        in the passed address, but will likely be different if the port number
-        contained in the passed address was zero.
+      integer:
+      An integer port on which server will accept RPC requests.
     """
         raise NotImplementedError()
 
     @abc.abstractmethod
     def start(self):
-        """Starts this Server's service of RPCs.
+        """Starts this Server.
 
-    This method may only be called while the server is not serving RPCs (i.e. it
-    is not idempotent).
+    This method may only be called once. (i.e. it is not idempotent).
     """
         raise NotImplementedError()
 
     @abc.abstractmethod
     def stop(self, grace):
-        """Stops this Server's service of RPCs.
+        """Stops this Server.
 
-    All calls to this method immediately stop service of new RPCs. When existing
-    RPCs are aborted is controlled by the grace period parameter passed to this
-    method.
+    This method immediately stop service of new RPCs in all cases.
+    If a grace period is specified, this method returns immediately
+    and all RPCs active at the end of the grace period are aborted.
 
-    This method may be called at any time and is idempotent. Passing a smaller
-    grace value than has been passed in a previous call will have the effect of
-    stopping the Server sooner. Passing a larger grace value than has been
-    passed in a previous call will not have the effect of stopping the server
-    later.
+    If a grace period is not specified, then all existing RPCs are
+    teriminated immediately and the this method blocks until the last
+    RPC handler terminates.
 
-    This method does not block for any significant length of time. If None is
-    passed as the grace value, existing RPCs are immediately aborted and this
-    method blocks until this Server is completely stopped.
+    This method is idempotent and may be called at any time. Passing a smaller
+    grace value in subsequentcall will have the effect of stopping the Server
+    sooner. Passing a larger grace value in subsequent call *will not* have the
+    effect of stopping the server later (i.e. the most restrictive grace
+    value is used).
 
     Args:
-      grace: A duration of time in seconds or None. If a duration of time in
-        seconds, the time to allow existing RPCs to complete before being
-        aborted by this Server's stopping. If None, all RPCs will be aborted
-        immediately and this method will block until this Server is completely
-        stopped.
+      grace: A duration of time in seconds or None.
 
     Returns:
       A threading.Event that will be set when this Server has completely
-      stopped. The returned event may not be set until after the full grace
-      period (if some ongoing RPC continues for the full length of the period)
-      of it may be set much sooner (such as if this Server had no RPCs underway
-      at the time it was stopped or if all RPCs that it had underway completed
-      very early in the grace period).
+      stopped, i.e. when running RPCs either complete or are aborted and
+      all handlers have terminated.
     """
         raise NotImplementedError()
 
@@ -995,14 +1004,13 @@ def unary_unary_rpc_method_handler(behavior,
     """Creates an RpcMethodHandler for a unary-unary RPC method.
 
   Args:
-    behavior: The implementation of an RPC method as a callable behavior taking
-      a single request value and returning a single response value.
-    request_deserializer: An optional request deserialization behavior.
-    response_serializer: An optional response serialization behavior.
+    behavior: The implementation of an RPC that accepts one request and returns
+    one response.
+    request_deserializer: An optional behavior for request deserialization.
+    response_serializer: An optional behavior for response serialization.
 
   Returns:
-    An RpcMethodHandler for a unary-unary RPC method constructed from the given
-      parameters.
+    An RpcMethodHandler object that is typically used by grpc.Server.
   """
     from grpc import _utilities  # pylint: disable=cyclic-import
     return _utilities.RpcMethodHandler(False, False, request_deserializer,
@@ -1016,14 +1024,13 @@ def unary_stream_rpc_method_handler(behavior,
     """Creates an RpcMethodHandler for a unary-stream RPC method.
 
   Args:
-    behavior: The implementation of an RPC method as a callable behavior taking
-      a single request value and returning an iterator of response values.
-    request_deserializer: An optional request deserialization behavior.
-    response_serializer: An optional response serialization behavior.
+    behavior: The implementation of an RPC that accepts one request and returns
+      an iterator of response values.
+    request_deserializer: An optional behavior for request deserialization.
+    response_serializer: An optional behavior for response serialization.
 
   Returns:
-    An RpcMethodHandler for a unary-stream RPC method constructed from the
-      given parameters.
+    An RpcMethodHandler object that is typically used by grpc.Server.
   """
     from grpc import _utilities  # pylint: disable=cyclic-import
     return _utilities.RpcMethodHandler(False, True, request_deserializer,
@@ -1037,14 +1044,13 @@ def stream_unary_rpc_method_handler(behavior,
     """Creates an RpcMethodHandler for a stream-unary RPC method.
 
   Args:
-    behavior: The implementation of an RPC method as a callable behavior taking
-      an iterator of request values and returning a single response value.
-    request_deserializer: An optional request deserialization behavior.
-    response_serializer: An optional response serialization behavior.
+    behavior: The implementation of an RPC that accepts an iterator of request
+    values and returns a single response value.
+    request_deserializer: An optional behavior for request deserialization.
+    response_serializer: An optional behavior for response serialization.
 
   Returns:
-    An RpcMethodHandler for a stream-unary RPC method constructed from the
-      given parameters.
+    An RpcMethodHandler object that is typically used by grpc.Server.
   """
     from grpc import _utilities  # pylint: disable=cyclic-import
     return _utilities.RpcMethodHandler(True, False, request_deserializer,
@@ -1058,15 +1064,13 @@ def stream_stream_rpc_method_handler(behavior,
     """Creates an RpcMethodHandler for a stream-stream RPC method.
 
   Args:
-    behavior: The implementation of an RPC method as a callable behavior taking
-      an iterator of request values and returning an iterator of response
-      values.
-    request_deserializer: An optional request deserialization behavior.
-    response_serializer: An optional response serialization behavior.
+    behavior: The implementation of an RPC that accepts an iterator of request
+    values and returns an iterator of response values.
+    request_deserializer: An optional behavior for request deserialization.
+    response_serializer: An optional behavior for response serialization.
 
   Returns:
-    An RpcMethodHandler for a stream-stream RPC method constructed from the
-      given parameters.
+    An RpcMethodHandler object that is typically used by grpc.Server.
   """
     from grpc import _utilities  # pylint: disable=cyclic-import
     return _utilities.RpcMethodHandler(True, True, request_deserializer,
@@ -1075,15 +1079,16 @@ def stream_stream_rpc_method_handler(behavior,
 
 
 def method_handlers_generic_handler(service, method_handlers):
-    """Creates a grpc.GenericRpcHandler from RpcMethodHandlers.
+    """Creates a GenericRpcHandler from RpcMethodHandlers.
 
   Args:
-    service: A service name to be used for the given method handlers.
-    method_handlers: A dictionary from method name to RpcMethodHandler
-      implementing the named method.
+    service: The name of the service that is implemented by the method_handlers.
+    method_handlers: A dictionary that maps method names to corresponding
+    RpcMethodHandler.
 
   Returns:
-    A GenericRpcHandler constructed from the given parameters.
+    A GenericRpcHandler. This is typically added to the grpc.Server object
+    with add_generic_rpc_handlers() before starting the server.
   """
     from grpc import _utilities  # pylint: disable=cyclic-import
     return _utilities.DictionaryGenericHandler(service, method_handlers)
@@ -1095,12 +1100,12 @@ def ssl_channel_credentials(root_certificates=None,
     """Creates a ChannelCredentials for use with an SSL-enabled Channel.
 
   Args:
-    root_certificates: The PEM-encoded root certificates or unset to ask for
-      them to be retrieved from a default location.
-    private_key: The PEM-encoded private key to use or unset if no private key
-      should be used.
-    certificate_chain: The PEM-encoded certificate chain to use or unset if no
-      certificate chain should be used.
+    root_certificates: The PEM-encoded root certificates as a byte string,
+    or None to retrieve them from a default location chosen by gRPC runtime.
+    private_key: The PEM-encoded private key as a byte string, or None if no
+    private key should be used.
+    certificate_chain: The PEM-encoded certificate chain as a byte string
+    to use or or None if no certificate chain should be used.
 
   Returns:
     A ChannelCredentials for use with an SSL-enabled Channel.
@@ -1117,9 +1122,8 @@ def metadata_call_credentials(metadata_plugin, name=None):
     """Construct CallCredentials from an AuthMetadataPlugin.
 
   Args:
-    metadata_plugin: An AuthMetadataPlugin to use as the authentication behavior
-      in the created CallCredentials.
-    name: A name for the plugin.
+    metadata_plugin: An AuthMetadataPlugin to use for authentication.
+    name: An optional name for the plugin.
 
   Returns:
     A CallCredentials.
@@ -1142,7 +1146,8 @@ def access_token_call_credentials(access_token):
 
   Args:
     access_token: A string to place directly in the http request
-      authorization header, ie "authorization: Bearer <access_token>".
+      authorization header, for example
+      "authorization: Bearer <access_token>".
 
   Returns:
     A CallCredentials.
@@ -1173,12 +1178,12 @@ def composite_channel_credentials(channel_credentials, *call_credentials):
     """Compose a ChannelCredentials and one or more CallCredentials objects.
 
   Args:
-    channel_credentials: A ChannelCredentials.
+    channel_credentials: A ChannelCredentials object.
     *call_credentials: One or more CallCredentials objects.
 
   Returns:
     A ChannelCredentials composed of the given ChannelCredentials and
-      CallCredentials objects.
+    CallCredentials objects.
   """
     from grpc import _credential_composition  # pylint: disable=cyclic-import
     cygrpc_call_credentials = tuple(
@@ -1195,18 +1200,18 @@ def ssl_server_credentials(private_key_certificate_chain_pairs,
     """Creates a ServerCredentials for use with an SSL-enabled Server.
 
   Args:
-    private_key_certificate_chain_pairs: A nonempty sequence each element of
-      which is a pair the first element of which is a PEM-encoded private key
-      and the second element of which is the corresponding PEM-encoded
-      certificate chain.
-    root_certificates: PEM-encoded client root certificates to be used for
-      verifying authenticated clients. If omitted, require_client_auth must also
-      be omitted or be False.
-    require_client_auth: A boolean indicating whether or not to require clients
-      to be authenticated. May only be True if root_certificates is not None.
+    private_key_certificate_chain_pairs: A list of pairs of the form
+      [PEM-encoded private key, PEM-encoded certificate chain].
+    root_certificates: An optional byte string of PEM-encoded client root
+      certificates that the server will use to verify client authentication.
+      If omitted, require_client_auth must also be False.
+    require_client_auth: A boolean indicating whether or not to require
+      clients to be authenticated. May only be True if root_certificates
+      is not None.
 
   Returns:
-    A ServerCredentials for use with an SSL-enabled Server.
+    A ServerCredentials for use with an SSL-enabled Server. Typically, this
+    object is an argument to add_secure_port() method during server setup.
   """
     if len(private_key_certificate_chain_pairs) == 0:
         raise ValueError(
@@ -1224,18 +1229,17 @@ def ssl_server_credentials(private_key_certificate_chain_pairs,
 
 
 def channel_ready_future(channel):
-    """Creates a Future tracking when a Channel is ready.
+    """Creates a Future that tracks when a Channel is ready.
 
-  Cancelling the returned Future does not tell the given Channel to abandon
-  attempts it may have been making to connect; cancelling merely deactivates the
-  returned Future's subscription to the given Channel's connectivity.
+  Cancelling the Future does not affect the channel's state machine.
+  It merely decouples the Future from channel state machine.
 
   Args:
-    channel: A Channel.
+    channel: A Channel object.
 
   Returns:
-    A Future that matures when the given Channel has connectivity
-      ChannelConnectivity.READY.
+    A Future object that matures when the channel connectivity is
+    ChannelConnectivity.READY.
   """
     from grpc import _utilities  # pylint: disable=cyclic-import
     return _utilities.channel_ready_future(channel)
@@ -1245,12 +1249,12 @@ def insecure_channel(target, options=None):
     """Creates an insecure Channel to a server.
 
   Args:
-    target: The target to which to connect.
-    options: A sequence of string-value pairs according to which to configure
-      the created channel.
+    target: The server address
+    options: An optional list of key-value pairs (channel args in gRPC runtime)
+    to configure the channel.
 
   Returns:
-    A Channel to the target through which RPCs may be conducted.
+    A Channel object.
   """
     from grpc import _channel  # pylint: disable=cyclic-import
     return _channel.Channel(target, () if options is None else options, None)
@@ -1260,13 +1264,13 @@ def secure_channel(target, credentials, options=None):
     """Creates a secure Channel to a server.
 
   Args:
-    target: The target to which to connect.
+    target: The server address.
     credentials: A ChannelCredentials instance.
-    options: A sequence of string-value pairs according to which to configure
-      the created channel.
+    options: An optional list of key-value pairs (channel args in gRPC runtime)
+    to configure the channel.
 
   Returns:
-    A Channel to the target through which RPCs may be conducted.
+    A Channel object.
   """
     from grpc import _channel  # pylint: disable=cyclic-import
     return _channel.Channel(target, () if options is None else options,
@@ -1280,21 +1284,19 @@ def server(thread_pool,
     """Creates a Server with which RPCs can be serviced.
 
   Args:
-    thread_pool: A futures.ThreadPoolExecutor to be used by the returned Server
-      to service RPCs.
-    handlers: An optional sequence of GenericRpcHandlers to be used to service
-      RPCs after the returned Server is started. These handlers need not be the
-      only handlers the server will use to service RPCs; other handlers may
-      later be added by calling add_generic_rpc_handlers any time before the
-      returned Server is started.
-    options: A sequence of string-value pairs according to which to configure
-      the created server.
+    thread_pool: A futures.ThreadPoolExecutor to be used by the Server
+      to execute RPC handlers.
+    handlers: An optional list of GenericRpcHandlers used for executing RPCs.
+      More handlers may be added by calling add_generic_rpc_handlers any time
+      before the server is started.
+    options: An optional list of key-value pairs (channel args in gRPC runtime)
+    to configure the channel.
     maximum_concurrent_rpcs: The maximum number of concurrent RPCs this server
-      will service before returning status RESOURCE_EXHAUSTED, or None to
+      will service before returning RESOURCE_EXHAUSTED status, or None to
       indicate no limit.
 
   Returns:
-    A Server with which RPCs can be serviced.
+    A Server object.
   """
     from grpc import _server  # pylint: disable=cyclic-import
     return _server.Server(thread_pool, () if handlers is None else handlers, ()

test/core/transport/bdp_estimator_test.c

@@ -33,6 +33,7 @@
 
 #include "src/core/lib/transport/bdp_estimator.h"
 
+#include <grpc/grpc.h>
 #include <grpc/support/alloc.h>
 #include <grpc/support/log.h>
 #include <grpc/support/string_util.h>
@@ -64,6 +65,8 @@ static void add_samples(grpc_bdp_estimator *estimator, int64_t *samples,
     GPR_ASSERT(grpc_bdp_estimator_add_incoming_bytes(estimator, samples[i]) ==
                false);
   }
+  gpr_sleep_until(gpr_time_add(gpr_now(GPR_CLOCK_REALTIME),
+                               gpr_time_from_millis(1, GPR_TIMESPAN)));
   grpc_bdp_estimator_complete_ping(estimator);
 }
 
@@ -123,24 +126,25 @@ static void test_get_estimate_random_values(size_t n) {
   gpr_log(GPR_INFO, "test_get_estimate_random_values(%" PRIdPTR ")", n);
   grpc_bdp_estimator est;
   grpc_bdp_estimator_init(&est, "test");
-  int min = INT_MAX;
-  int max = 65535;  // Windows rand() has limited range, make sure the ASSERT
-                    // passes
+  const int kMaxSample = 65535;
+  int min = kMaxSample;
+  int max = 0;
   for (size_t i = 0; i < n; i++) {
-    int sample = rand();
+    int sample = rand() % (kMaxSample + 1);
     if (sample < min) min = sample;
     if (sample > max) max = sample;
     add_sample(&est, sample);
     if (i >= 3) {
       gpr_log(GPR_DEBUG, "est:%" PRId64 " min:%d max:%d", get_estimate(&est),
               min, max);
-      GPR_ASSERT(get_estimate(&est) <= 2 * next_pow_2(max));
+      GPR_ASSERT(get_estimate(&est) <= GPR_MAX(65536, 2 * next_pow_2(max)));
     }
   }
 }
 
 int main(int argc, char **argv) {
   grpc_test_init(argc, argv);
+  grpc_init();
   test_noop();
   test_get_estimate_no_samples();
   test_get_estimate_1_sample();
@@ -149,5 +153,6 @@ int main(int argc, char **argv) {
   for (size_t i = 3; i < 1000; i = i * 3 / 2) {
     test_get_estimate_random_values(i);
   }
+  grpc_shutdown();
   return 0;
 }

test/core/util/trickle_endpoint.c

@@ -44,6 +44,8 @@
 #include <grpc/support/useful.h>
 #include "src/core/lib/slice/slice_internal.h"
 
+#define WRITE_BUFFER_SIZE (2 * 1024 * 1024)
+
 typedef struct {
   grpc_endpoint base;
   double bytes_per_second;
@@ -55,6 +57,7 @@ typedef struct {
   grpc_slice_buffer writing_buffer;
   grpc_error *error;
   bool writing;
+  grpc_closure *write_cb;
 } trickle_endpoint;
 
 static void te_read(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep,
@@ -63,10 +66,20 @@ static void te_read(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep,
   grpc_endpoint_read(exec_ctx, te->wrapped, slices, cb);
 }
 
+static void maybe_call_write_cb_locked(grpc_exec_ctx *exec_ctx,
+                                       trickle_endpoint *te) {
+  if (te->write_cb != NULL && (te->error != GRPC_ERROR_NONE ||
+                               te->write_buffer.length <= WRITE_BUFFER_SIZE)) {
+    grpc_closure_sched(exec_ctx, te->write_cb, GRPC_ERROR_REF(te->error));
+    te->write_cb = NULL;
+  }
+}
+
 static void te_write(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep,
                      grpc_slice_buffer *slices, grpc_closure *cb) {
   trickle_endpoint *te = (trickle_endpoint *)ep;
   gpr_mu_lock(&te->mu);
+  GPR_ASSERT(te->write_cb == NULL);
   if (te->write_buffer.length == 0) {
     te->last_write = gpr_now(GPR_CLOCK_MONOTONIC);
   }
@@ -74,7 +87,8 @@ static void te_write(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep,
     grpc_slice_buffer_add(&te->write_buffer,
                           grpc_slice_copy(slices->slices[i]));
   }
-  grpc_closure_sched(exec_ctx, cb, GRPC_ERROR_REF(te->error));
+  te->write_cb = cb;
+  maybe_call_write_cb_locked(exec_ctx, te);
   gpr_mu_unlock(&te->mu);
 }
 
@@ -102,6 +116,7 @@ static void te_shutdown(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep,
   if (te->error == GRPC_ERROR_NONE) {
     te->error = GRPC_ERROR_REF(why);
   }
+  maybe_call_write_cb_locked(exec_ctx, te);
   gpr_mu_unlock(&te->mu);
   grpc_endpoint_shutdown(exec_ctx, te->wrapped, why);
 }
@@ -157,6 +172,7 @@ grpc_endpoint *grpc_trickle_endpoint_create(grpc_endpoint *wrap,
   te->base.vtable = &vtable;
   te->wrapped = wrap;
   te->bytes_per_second = bytes_per_second;
+  te->write_cb = NULL;
   gpr_mu_init(&te->mu);
   grpc_slice_buffer_init(&te->write_buffer);
   grpc_slice_buffer_init(&te->writing_buffer);
@@ -187,9 +203,18 @@ size_t grpc_trickle_endpoint_trickle(grpc_exec_ctx *exec_ctx,
       grpc_endpoint_write(
           exec_ctx, te->wrapped, &te->writing_buffer,
           grpc_closure_create(te_finish_write, te, grpc_schedule_on_exec_ctx));
+      maybe_call_write_cb_locked(exec_ctx, te);
     }
   }
   size_t backlog = te->write_buffer.length;
   gpr_mu_unlock(&te->mu);
   return backlog;
 }
+
+size_t grpc_trickle_get_backlog(grpc_endpoint *ep) {
+  trickle_endpoint *te = (trickle_endpoint *)ep;
+  gpr_mu_lock(&te->mu);
+  size_t backlog = te->write_buffer.length;
+  gpr_mu_unlock(&te->mu);
+  return backlog;
+}

test/core/util/trickle_endpoint.h

@@ -43,4 +43,6 @@ grpc_endpoint *grpc_trickle_endpoint_create(grpc_endpoint *wrap,
 size_t grpc_trickle_endpoint_trickle(grpc_exec_ctx *exec_ctx,
                                      grpc_endpoint *endpoint);
 
+size_t grpc_trickle_get_backlog(grpc_endpoint *endpoint);
+
 #endif

test/cpp/microbenchmarks/BUILD

@@ -92,7 +92,7 @@ cc_test(
 cc_test(
     name = "bm_fullstack_trickle",
     srcs = ["bm_fullstack_trickle.cc"],
-    deps = [":helpers"],
+    deps = [":helpers", "//external:gflags"],
 )
 
 cc_test(

test/cpp/microbenchmarks/bm_fullstack_trickle.cc

@@ -34,6 +34,8 @@
 /* Benchmark gRPC end2end in various configurations */
 
 #include <benchmark/benchmark.h>
+#include <gflags/gflags.h>
+#include <fstream>
 #include "src/core/lib/profiling/timers.h"
 #include "src/cpp/client/create_channel_internal.h"
 #include "src/proto/grpc/testing/echo.grpc.pb.h"
@@ -45,16 +47,57 @@ extern "C" {
 #include "test/core/util/trickle_endpoint.h"
 }
 
+DEFINE_bool(log, false, "Log state to CSV files");
+DEFINE_int32(
+    warmup_megabytes, 1,
+    "Number of megabytes to pump before collecting flow control stats");
+DEFINE_int32(
+    warmup_iterations, 100,
+    "Number of megabytes to pump before collecting flow control stats");
+DEFINE_int32(warmup_max_time_seconds, 10,
+             "Maximum number of seconds to run warmup loop");
+
 namespace grpc {
 namespace testing {
 
 static void* tag(intptr_t x) { return reinterpret_cast<void*>(x); }
 
+template <class A0>
+static void write_csv(std::ostream* out, A0&& a0) {
+  if (!out) return;
+  (*out) << a0 << "\n";
+}
+
+template <class A0, class... Arg>
+static void write_csv(std::ostream* out, A0&& a0, Arg&&... arg) {
+  if (!out) return;
+  (*out) << a0 << ",";
+  write_csv(out, std::forward<Arg>(arg)...);
+}
+
 class TrickledCHTTP2 : public EndpointPairFixture {
  public:
-  TrickledCHTTP2(Service* service, size_t megabits_per_second)
-      : EndpointPairFixture(service, MakeEndpoints(megabits_per_second),
-                            FixtureConfiguration()) {}
+  TrickledCHTTP2(Service* service, size_t message_size,
+                 size_t kilobits_per_second)
+      : EndpointPairFixture(service, MakeEndpoints(kilobits_per_second),
+                            FixtureConfiguration()) {
+    if (FLAGS_log) {
+      std::ostringstream fn;
+      fn << "trickle." << message_size << "." << kilobits_per_second << ".csv";
+      log_.reset(new std::ofstream(fn.str().c_str()));
+      write_csv(log_.get(), "t", "iteration", "client_backlog",
+                "server_backlog", "client_t_stall", "client_s_stall",
+                "server_t_stall", "server_s_stall", "client_t_outgoing",
+                "server_t_outgoing", "client_t_incoming", "server_t_incoming",
+                "client_s_outgoing_delta", "server_s_outgoing_delta",
+                "client_s_incoming_delta", "server_s_incoming_delta",
+                "client_s_announce_window", "server_s_announce_window",
+                "client_peer_iws", "client_local_iws", "client_sent_iws",
+                "client_acked_iws", "server_peer_iws", "server_local_iws",
+                "server_sent_iws", "server_acked_iws", "client_queued_bytes",
+                "server_queued_bytes");
+    }
+  }
 
   void AddToLabel(std::ostream& out, benchmark::State& state) {
     out << " writes/iter:"
@@ -75,7 +118,58 @@ class TrickledCHTTP2 : public EndpointPairFixture {
             (double)state.iterations());
   }
 
-  void Step() {
+  void Log(int64_t iteration) {
+    auto now = gpr_time_sub(gpr_now(GPR_CLOCK_MONOTONIC), start_);
+    grpc_chttp2_transport* client =
+        reinterpret_cast<grpc_chttp2_transport*>(client_transport_);
+    grpc_chttp2_transport* server =
+        reinterpret_cast<grpc_chttp2_transport*>(server_transport_);
+    grpc_chttp2_stream* client_stream =
+        client->stream_map.count == 1
+            ? static_cast<grpc_chttp2_stream*>(client->stream_map.values[0])
+            : nullptr;
+    grpc_chttp2_stream* server_stream =
+        server->stream_map.count == 1
+            ? static_cast<grpc_chttp2_stream*>(server->stream_map.values[0])
+            : nullptr;
+    write_csv(
+        log_.get(), static_cast<double>(now.tv_sec) +
+                        1e-9 * static_cast<double>(now.tv_nsec),
+        iteration, grpc_trickle_get_backlog(endpoint_pair_.client),
+        grpc_trickle_get_backlog(endpoint_pair_.server),
+        client->lists[GRPC_CHTTP2_LIST_STALLED_BY_TRANSPORT].head != nullptr,
+        client->lists[GRPC_CHTTP2_LIST_STALLED_BY_STREAM].head != nullptr,
+        server->lists[GRPC_CHTTP2_LIST_STALLED_BY_TRANSPORT].head != nullptr,
+        server->lists[GRPC_CHTTP2_LIST_STALLED_BY_STREAM].head != nullptr,
+        client->outgoing_window, server->outgoing_window,
+        client->incoming_window, server->incoming_window,
+        client_stream ? client_stream->outgoing_window_delta : -1,
+        server_stream ? server_stream->outgoing_window_delta : -1,
+        client_stream ? client_stream->incoming_window_delta : -1,
+        server_stream ? server_stream->incoming_window_delta : -1,
+        client_stream ? client_stream->announce_window : -1,
+        server_stream ? server_stream->announce_window : -1,
+        client->settings[GRPC_PEER_SETTINGS]
+                        [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE],
+        client->settings[GRPC_LOCAL_SETTINGS]
+                        [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE],
+        client->settings[GRPC_SENT_SETTINGS]
+                        [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE],
+        client->settings[GRPC_ACKED_SETTINGS]
+                        [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE],
+        server->settings[GRPC_PEER_SETTINGS]
+                        [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE],
+        server->settings[GRPC_LOCAL_SETTINGS]
+                        [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE],
+        server->settings[GRPC_SENT_SETTINGS]
+                        [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE],
+        server->settings[GRPC_ACKED_SETTINGS]
+                        [GRPC_CHTTP2_SETTINGS_INITIAL_WINDOW_SIZE],
+        client_stream ? client_stream->flow_controlled_buffer.length : 0,
+        server_stream ? server_stream->flow_controlled_buffer.length : 0);
+  }
+
+  void Step(bool update_stats) {
     grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT;
     size_t client_backlog =
         grpc_trickle_endpoint_trickle(&exec_ctx, endpoint_pair_.client);
@@ -83,10 +177,12 @@ class TrickledCHTTP2 : public EndpointPairFixture {
         grpc_trickle_endpoint_trickle(&exec_ctx, endpoint_pair_.server);
     grpc_exec_ctx_finish(&exec_ctx);
 
-    UpdateStats((grpc_chttp2_transport*)client_transport_, &client_stats_,
-                client_backlog);
-    UpdateStats((grpc_chttp2_transport*)server_transport_, &server_stats_,
-                server_backlog);
+    if (update_stats) {
+      UpdateStats((grpc_chttp2_transport*)client_transport_, &client_stats_,
+                  client_backlog);
+      UpdateStats((grpc_chttp2_transport*)server_transport_, &server_stats_,
+                  server_backlog);
+    }
   }
 
  private:
@@ -97,6 +193,8 @@ class TrickledCHTTP2 : public EndpointPairFixture {
   };
   Stats client_stats_;
   Stats server_stats_;
+  std::unique_ptr<std::ofstream> log_;
+  gpr_timespec start_ = gpr_now(GPR_CLOCK_MONOTONIC);
 
   grpc_endpoint_pair MakeEndpoints(size_t kilobits) {
     grpc_endpoint_pair p;
@@ -123,13 +221,15 @@ class TrickledCHTTP2 : public EndpointPairFixture {
 // force library initialization
 auto& force_library_initialization = Library::get();
 
-static void TrickleCQNext(TrickledCHTTP2* fixture, void** t, bool* ok) {
+static void TrickleCQNext(TrickledCHTTP2* fixture, void** t, bool* ok,
+                          int64_t iteration) {
   while (true) {
+    fixture->Log(iteration);
     switch (fixture->cq()->AsyncNext(
         t, ok, gpr_time_add(gpr_now(GPR_CLOCK_MONOTONIC),
                             gpr_time_from_micros(100, GPR_TIMESPAN)))) {
       case CompletionQueue::TIMEOUT:
-        fixture->Step();
+        fixture->Step(iteration != -1);
         break;
       case CompletionQueue::SHUTDOWN:
         GPR_ASSERT(false);
@@ -143,7 +243,7 @@ static void TrickleCQNext(TrickledCHTTP2* fixture, void** t, bool* ok) {
 static void BM_PumpStreamServerToClient_Trickle(benchmark::State& state) {
   EchoTestService::AsyncService service;
   std::unique_ptr<TrickledCHTTP2> fixture(
-      new TrickledCHTTP2(&service, state.range(1)));
+      new TrickledCHTTP2(&service, state.range(0), state.range(1)));
   {
     EchoResponse send_response;
     EchoResponse recv_response;
@@ -163,18 +263,19 @@ static void BM_PumpStreamServerToClient_Trickle(benchmark::State& state) {
     void* t;
     bool ok;
     while (need_tags) {
-      TrickleCQNext(fixture.get(), &t, &ok);
+      TrickleCQNext(fixture.get(), &t, &ok, -1);
       GPR_ASSERT(ok);
       int i = (int)(intptr_t)t;
       GPR_ASSERT(need_tags & (1 << i));
       need_tags &= ~(1 << i);
     }
     request_rw->Read(&recv_response, tag(0));
-    while (state.KeepRunning()) {
+    auto inner_loop = [&](bool in_warmup) {
       GPR_TIMER_SCOPE("BenchmarkCycle", 0);
       response_rw.Write(send_response, tag(1));
       while (true) {
-        TrickleCQNext(fixture.get(), &t, &ok);
+        TrickleCQNext(fixture.get(), &t, &ok,
+                      in_warmup ? -1 : state.iterations());
         if (t == tag(0)) {
           request_rw->Read(&recv_response, tag(0));
         } else if (t == tag(1)) {
@@ -183,11 +284,26 @@ static void BM_PumpStreamServerToClient_Trickle(benchmark::State& state) {
           GPR_ASSERT(false);
         }
       }
+    };
+    gpr_timespec warmup_start = gpr_now(GPR_CLOCK_MONOTONIC);
+    for (int i = 0;
+         i < GPR_MAX(FLAGS_warmup_iterations, FLAGS_warmup_megabytes * 1024 *
+                                                  1024 / (14 + state.range(0)));
+         i++) {
+      inner_loop(true);
+      if (gpr_time_cmp(gpr_time_sub(gpr_now(GPR_CLOCK_MONOTONIC), warmup_start),
+                       gpr_time_from_seconds(FLAGS_warmup_max_time_seconds,
+                                             GPR_TIMESPAN)) > 0) {
+        break;
+      }
+    }
+    while (state.KeepRunning()) {
+      inner_loop(false);
     }
     response_rw.Finish(Status::OK, tag(1));
     need_tags = (1 << 0) | (1 << 1);
     while (need_tags) {
-      TrickleCQNext(fixture.get(), &t, &ok);
+      TrickleCQNext(fixture.get(), &t, &ok, -1);
       int i = (int)(intptr_t)t;
       GPR_ASSERT(need_tags & (1 << i));
       need_tags &= ~(1 << i);
@@ -204,10 +320,10 @@ static void BM_PumpStreamServerToClient_Trickle(benchmark::State& state) {
 
 static void TrickleArgs(benchmark::internal::Benchmark* b) {
   for (int i = 1; i <= 128 * 1024 * 1024; i *= 8) {
-    for (int j = 1; j <= 128 * 1024 * 1024; j *= 8) {
+    for (int j = 64; j <= 128 * 1024 * 1024; j *= 8) {
       double expected_time =
           static_cast<double>(14 + i) / (125.0 * static_cast<double>(j));
-      if (expected_time > 0.01) continue;
+      if (expected_time > 2.0) continue;
       b->Args({i, j});
     }
   }
@@ -217,4 +333,8 @@ BENCHMARK(BM_PumpStreamServerToClient_Trickle)->Apply(TrickleArgs);
 }
 }
 
-BENCHMARK_MAIN();
+int main(int argc, char** argv) {
+  ::benchmark::Initialize(&argc, argv);
+  ::google::ParseCommandLineFlags(&argc, &argv, false);
+  ::benchmark::RunSpecifiedBenchmarks();
+}

test/cpp/qps/client.h

@@ -443,11 +443,8 @@ class ClientImpl : public Client {
       create_stub_;
 };
 
-std::unique_ptr<Client> CreateSynchronousUnaryClient(const ClientConfig& args);
-std::unique_ptr<Client> CreateSynchronousStreamingClient(
-    const ClientConfig& args);
-std::unique_ptr<Client> CreateAsyncUnaryClient(const ClientConfig& args);
-std::unique_ptr<Client> CreateAsyncStreamingClient(const ClientConfig& args);
+std::unique_ptr<Client> CreateSynchronousClient(const ClientConfig& args);
+std::unique_ptr<Client> CreateAsyncClient(const ClientConfig& args);
 std::unique_ptr<Client> CreateGenericAsyncStreamingClient(
     const ClientConfig& args);
 

test/cpp/qps/client_async.cc

@@ -301,9 +301,9 @@ class AsyncUnaryClient final
 };
 
 template <class RequestType, class ResponseType>
-class ClientRpcContextStreamingImpl : public ClientRpcContext {
+class ClientRpcContextStreamingPingPongImpl : public ClientRpcContext {
  public:
-  ClientRpcContextStreamingImpl(
+  ClientRpcContextStreamingPingPongImpl(
       BenchmarkService::Stub* stub, const RequestType& req,
       std::function<gpr_timespec()> next_issue,
       std::function<std::unique_ptr<
@@ -321,7 +321,7 @@ class ClientRpcContextStreamingImpl : public ClientRpcContext {
         callback_(on_done),
         next_issue_(next_issue),
         start_req_(start_req) {}
-  ~ClientRpcContextStreamingImpl() override {}
+  ~ClientRpcContextStreamingPingPongImpl() override {}
   void Start(CompletionQueue* cq, const ClientConfig& config) override {
     StartInternal(cq, config.messages_per_stream());
   }
@@ -382,8 +382,8 @@ class ClientRpcContextStreamingImpl : public ClientRpcContext {
     }
   }
   void StartNewClone(CompletionQueue* cq) override {
-    auto* clone = new ClientRpcContextStreamingImpl(stub_, req_, next_issue_,
-                                                    start_req_, callback_);
+    auto* clone = new ClientRpcContextStreamingPingPongImpl(
+        stub_, req_, next_issue_, start_req_, callback_);
     clone->StartInternal(cq, messages_per_stream_);
   }
 
@@ -422,23 +422,23 @@ class ClientRpcContextStreamingImpl : public ClientRpcContext {
 
   void StartInternal(CompletionQueue* cq, int messages_per_stream) {
     cq_ = cq;
-    next_state_ = State::STREAM_IDLE;
-    stream_ = start_req_(stub_, &context_, cq, ClientRpcContext::tag(this));
     messages_per_stream_ = messages_per_stream;
     messages_issued_ = 0;
+    next_state_ = State::STREAM_IDLE;
+    stream_ = start_req_(stub_, &context_, cq, ClientRpcContext::tag(this));
   }
 };
 
-class AsyncStreamingClient final
+class AsyncStreamingPingPongClient final
     : public AsyncClient<BenchmarkService::Stub, SimpleRequest> {
  public:
-  explicit AsyncStreamingClient(const ClientConfig& config)
+  explicit AsyncStreamingPingPongClient(const ClientConfig& config)
       : AsyncClient<BenchmarkService::Stub, SimpleRequest>(
             config, SetupCtx, BenchmarkStubCreator) {
     StartThreads(num_async_threads_);
   }
 
-  ~AsyncStreamingClient() override {}
+  ~AsyncStreamingPingPongClient() override {}
 
  private:
   static void CheckDone(grpc::Status s, SimpleResponse* response) {}
@@ -452,9 +452,250 @@ class AsyncStreamingClient final
   static ClientRpcContext* SetupCtx(BenchmarkService::Stub* stub,
                                     std::function<gpr_timespec()> next_issue,
                                     const SimpleRequest& req) {
-    return new ClientRpcContextStreamingImpl<SimpleRequest, SimpleResponse>(
-        stub, req, next_issue, AsyncStreamingClient::StartReq,
-        AsyncStreamingClient::CheckDone);
+    return new ClientRpcContextStreamingPingPongImpl<SimpleRequest,
+                                                     SimpleResponse>(
+        stub, req, next_issue, AsyncStreamingPingPongClient::StartReq,
+        AsyncStreamingPingPongClient::CheckDone);
+  }
+};
+
+template <class RequestType, class ResponseType>
+class ClientRpcContextStreamingFromClientImpl : public ClientRpcContext {
+ public:
+  ClientRpcContextStreamingFromClientImpl(
+      BenchmarkService::Stub* stub, const RequestType& req,
+      std::function<gpr_timespec()> next_issue,
+      std::function<std::unique_ptr<grpc::ClientAsyncWriter<RequestType>>(
+          BenchmarkService::Stub*, grpc::ClientContext*, ResponseType*,
+          CompletionQueue*, void*)>
+          start_req,
+      std::function<void(grpc::Status, ResponseType*)> on_done)
+      : context_(),
+        stub_(stub),
+        cq_(nullptr),
+        req_(req),
+        response_(),
+        next_state_(State::INVALID),
+        callback_(on_done),
+        next_issue_(next_issue),
+        start_req_(start_req) {}
+  ~ClientRpcContextStreamingFromClientImpl() override {}
+  void Start(CompletionQueue* cq, const ClientConfig& config) override {
+    StartInternal(cq);
+  }
+  bool RunNextState(bool ok, HistogramEntry* entry) override {
+    while (true) {
+      switch (next_state_) {
+        case State::STREAM_IDLE:
+          if (!next_issue_) {  // ready to issue
+            next_state_ = State::READY_TO_WRITE;
+          } else {
+            next_state_ = State::WAIT;
+          }
+          break;  // loop around, don't return
+        case State::WAIT:
+          alarm_.reset(
+              new Alarm(cq_, next_issue_(), ClientRpcContext::tag(this)));
+          next_state_ = State::READY_TO_WRITE;
+          return true;
+        case State::READY_TO_WRITE:
+          if (!ok) {
+            return false;
+          }
+          start_ = UsageTimer::Now();
+          next_state_ = State::WRITE_DONE;
+          stream_->Write(req_, ClientRpcContext::tag(this));
+          return true;
+        case State::WRITE_DONE:
+          if (!ok) {
+            return false;
+          }
+          entry->set_value((UsageTimer::Now() - start_) * 1e9);
+          next_state_ = State::STREAM_IDLE;
+          break;  // loop around
+        default:
+          GPR_ASSERT(false);
+          return false;
+      }
+    }
+  }
+  void StartNewClone(CompletionQueue* cq) override {
+    auto* clone = new ClientRpcContextStreamingFromClientImpl(
+        stub_, req_, next_issue_, start_req_, callback_);
+    clone->StartInternal(cq);
+  }
+
+ private:
+  grpc::ClientContext context_;
+  BenchmarkService::Stub* stub_;
+  CompletionQueue* cq_;
+  std::unique_ptr<Alarm> alarm_;
+  RequestType req_;
+  ResponseType response_;
+  enum State {
+    INVALID,
+    STREAM_IDLE,
+    WAIT,
+    READY_TO_WRITE,
+    WRITE_DONE,
+  };
+  State next_state_;
+  std::function<void(grpc::Status, ResponseType*)> callback_;
+  std::function<gpr_timespec()> next_issue_;
+  std::function<std::unique_ptr<grpc::ClientAsyncWriter<RequestType>>(
+      BenchmarkService::Stub*, grpc::ClientContext*, ResponseType*,
+      CompletionQueue*, void*)>
+      start_req_;
+  grpc::Status status_;
+  double start_;
+  std::unique_ptr<grpc::ClientAsyncWriter<RequestType>> stream_;
+
+  void StartInternal(CompletionQueue* cq) {
+    cq_ = cq;
+    stream_ = start_req_(stub_, &context_, &response_, cq,
+                         ClientRpcContext::tag(this));
+    next_state_ = State::STREAM_IDLE;
+  }
+};
+
+class AsyncStreamingFromClientClient final
+    : public AsyncClient<BenchmarkService::Stub, SimpleRequest> {
+ public:
+  explicit AsyncStreamingFromClientClient(const ClientConfig& config)
+      : AsyncClient<BenchmarkService::Stub, SimpleRequest>(
+            config, SetupCtx, BenchmarkStubCreator) {
+    StartThreads(num_async_threads_);
+  }
+
+  ~AsyncStreamingFromClientClient() override {}
+
+ private:
+  static void CheckDone(grpc::Status s, SimpleResponse* response) {}
+  static std::unique_ptr<grpc::ClientAsyncWriter<SimpleRequest>> StartReq(
+      BenchmarkService::Stub* stub, grpc::ClientContext* ctx,
+      SimpleResponse* resp, CompletionQueue* cq, void* tag) {
+    auto stream = stub->AsyncStreamingFromClient(ctx, resp, cq, tag);
+    return stream;
+  };
+  static ClientRpcContext* SetupCtx(BenchmarkService::Stub* stub,
+                                    std::function<gpr_timespec()> next_issue,
+                                    const SimpleRequest& req) {
+    return new ClientRpcContextStreamingFromClientImpl<SimpleRequest,
+                                                       SimpleResponse>(
+        stub, req, next_issue, AsyncStreamingFromClientClient::StartReq,
+        AsyncStreamingFromClientClient::CheckDone);
+  }
+};
+
+template <class RequestType, class ResponseType>
+class ClientRpcContextStreamingFromServerImpl : public ClientRpcContext {
+ public:
+  ClientRpcContextStreamingFromServerImpl(
+      BenchmarkService::Stub* stub, const RequestType& req,
+      std::function<gpr_timespec()> next_issue,
+      std::function<std::unique_ptr<grpc::ClientAsyncReader<ResponseType>>(
+          BenchmarkService::Stub*, grpc::ClientContext*, const RequestType&,
+          CompletionQueue*, void*)>
+          start_req,
+      std::function<void(grpc::Status, ResponseType*)> on_done)
+      : context_(),
+        stub_(stub),
+        cq_(nullptr),
+        req_(req),
+        response_(),
+        next_state_(State::INVALID),
+        callback_(on_done),
+        next_issue_(next_issue),
+        start_req_(start_req) {}
+  ~ClientRpcContextStreamingFromServerImpl() override {}
+  void Start(CompletionQueue* cq, const ClientConfig& config) override {
+    StartInternal(cq);
+  }
+  bool RunNextState(bool ok, HistogramEntry* entry) override {
+    while (true) {
+      switch (next_state_) {
+        case State::STREAM_IDLE:
+          if (!ok) {
+            return false;
+          }
+          start_ = UsageTimer::Now();
+          next_state_ = State::READ_DONE;
+          stream_->Read(&response_, ClientRpcContext::tag(this));
+          return true;
+        case State::READ_DONE:
+          if (!ok) {
+            return false;
+          }
+          entry->set_value((UsageTimer::Now() - start_) * 1e9);
+          callback_(status_, &response_);
+          next_state_ = State::STREAM_IDLE;
+          break;  // loop around
+        default:
+          GPR_ASSERT(false);
+          return false;
+      }
+    }
+  }
+  void StartNewClone(CompletionQueue* cq) override {
+    auto* clone = new ClientRpcContextStreamingFromServerImpl(
+        stub_, req_, next_issue_, start_req_, callback_);
+    clone->StartInternal(cq);
+  }
+
+ private:
+  grpc::ClientContext context_;
+  BenchmarkService::Stub* stub_;
+  CompletionQueue* cq_;
+  std::unique_ptr<Alarm> alarm_;
+  RequestType req_;
+  ResponseType response_;
+  enum State { INVALID, STREAM_IDLE, READ_DONE };
+  State next_state_;
+  std::function<void(grpc::Status, ResponseType*)> callback_;
+  std::function<gpr_timespec()> next_issue_;
+  std::function<std::unique_ptr<grpc::ClientAsyncReader<ResponseType>>(
+      BenchmarkService::Stub*, grpc::ClientContext*, const RequestType&,
+      CompletionQueue*, void*)>
+      start_req_;
+  grpc::Status status_;
+  double start_;
+  std::unique_ptr<grpc::ClientAsyncReader<ResponseType>> stream_;
+
+  void StartInternal(CompletionQueue* cq) {
+    // TODO(vjpai): Add support to rate-pace this
+    cq_ = cq;
+    next_state_ = State::STREAM_IDLE;
+    stream_ =
+        start_req_(stub_, &context_, req_, cq, ClientRpcContext::tag(this));
+  }
+};
+
+class AsyncStreamingFromServerClient final
+    : public AsyncClient<BenchmarkService::Stub, SimpleRequest> {
+ public:
+  explicit AsyncStreamingFromServerClient(const ClientConfig& config)
+      : AsyncClient<BenchmarkService::Stub, SimpleRequest>(
+            config, SetupCtx, BenchmarkStubCreator) {
+    StartThreads(num_async_threads_);
+  }
+
+  ~AsyncStreamingFromServerClient() override {}
+
+ private:
+  static void CheckDone(grpc::Status s, SimpleResponse* response) {}
+  static std::unique_ptr<grpc::ClientAsyncReader<SimpleResponse>> StartReq(
+      BenchmarkService::Stub* stub, grpc::ClientContext* ctx,
+      const SimpleRequest& req, CompletionQueue* cq, void* tag) {
+    auto stream = stub->AsyncStreamingFromServer(ctx, req, cq, tag);
+    return stream;
+  };
+  static ClientRpcContext* SetupCtx(BenchmarkService::Stub* stub,
+                                    std::function<gpr_timespec()> next_issue,
+                                    const SimpleRequest& req) {
+    return new ClientRpcContextStreamingFromServerImpl<SimpleRequest,
+                                                       SimpleResponse>(
+        stub, req, next_issue, AsyncStreamingFromServerClient::StartReq,
+        AsyncStreamingFromServerClient::CheckDone);
   }
 };
 
@@ -579,11 +820,11 @@ class ClientRpcContextGenericStreamingImpl : public ClientRpcContext {
     cq_ = cq;
     const grpc::string kMethodName(
         "/grpc.testing.BenchmarkService/StreamingCall");
+    messages_per_stream_ = messages_per_stream;
+    messages_issued_ = 0;
     next_state_ = State::STREAM_IDLE;
     stream_ = start_req_(stub_, &context_, kMethodName, cq,
                          ClientRpcContext::tag(this));
-    messages_per_stream_ = messages_per_stream;
-    messages_issued_ = 0;
   }
 };
 
@@ -620,11 +861,26 @@ class GenericAsyncStreamingClient final
   }
 };
 
-std::unique_ptr<Client> CreateAsyncUnaryClient(const ClientConfig& args) {
-  return std::unique_ptr<Client>(new AsyncUnaryClient(args));
-}
-std::unique_ptr<Client> CreateAsyncStreamingClient(const ClientConfig& args) {
-  return std::unique_ptr<Client>(new AsyncStreamingClient(args));
+std::unique_ptr<Client> CreateAsyncClient(const ClientConfig& config) {
+  switch (config.rpc_type()) {
+    case UNARY:
+      return std::unique_ptr<Client>(new AsyncUnaryClient(config));
+    case STREAMING:
+      return std::unique_ptr<Client>(new AsyncStreamingPingPongClient(config));
+    case STREAMING_FROM_CLIENT:
+      return std::unique_ptr<Client>(
+          new AsyncStreamingFromClientClient(config));
+    case STREAMING_FROM_SERVER:
+      return std::unique_ptr<Client>(
+          new AsyncStreamingFromServerClient(config));
+    case STREAMING_BOTH_WAYS:
+      // TODO(vjpai): Implement this
+      assert(false);
+      return nullptr;
+    default:
+      assert(false);
+      return nullptr;
+  }
 }
 std::unique_ptr<Client> CreateGenericAsyncStreamingClient(
     const ClientConfig& args) {

test/cpp/qps/client_sync.cc

@@ -137,7 +137,8 @@ class SynchronousUnaryClient final : public SynchronousClient {
   }
 };
 
-class SynchronousStreamingClient final : public SynchronousClient {
+template <class StreamType>
+class SynchronousStreamingClient : public SynchronousClient {
  public:
   SynchronousStreamingClient(const ClientConfig& config)
       : SynchronousClient(config),
@@ -145,30 +146,69 @@ class SynchronousStreamingClient final : public SynchronousClient {
         stream_(num_threads_),
         messages_per_stream_(config.messages_per_stream()),
         messages_issued_(num_threads_) {
+    StartThreads(num_threads_);
+  }
+  virtual ~SynchronousStreamingClient() {
+    std::vector<std::thread> cleanup_threads;
+    for (size_t i = 0; i < num_threads_; i++) {
+      cleanup_threads.emplace_back([this, i]() {
+        auto stream = &stream_[i];
+        if (*stream) {
+          // forcibly cancel the streams, then finish
+          context_[i].TryCancel();
+          (*stream)->Finish();
+          // don't log any error message on !ok since this was canceled
+        }
+      });
+    }
+    for (auto& th : cleanup_threads) {
+      th.join();
+    }
+  }
+
+ protected:
+  std::vector<grpc::ClientContext> context_;
+  std::vector<std::unique_ptr<StreamType>> stream_;
+  const int messages_per_stream_;
+  std::vector<int> messages_issued_;
+
+  void FinishStream(HistogramEntry* entry, size_t thread_idx) {
+    Status s = stream_[thread_idx]->Finish();
+    // don't set the value since the stream is failed and shouldn't be timed
+    entry->set_status(s.error_code());
+    if (!s.ok()) {
+      gpr_log(GPR_ERROR, "Stream %" PRIuPTR " received an error %s", thread_idx,
+              s.error_message().c_str());
+    }
+    context_[thread_idx].~ClientContext();
+    new (&context_[thread_idx]) ClientContext();
+  }
+};
+
+class SynchronousStreamingPingPongClient final
+    : public SynchronousStreamingClient<
+          grpc::ClientReaderWriter<SimpleRequest, SimpleResponse>> {
+ public:
+  SynchronousStreamingPingPongClient(const ClientConfig& config)
+      : SynchronousStreamingClient(config) {
     for (size_t thread_idx = 0; thread_idx < num_threads_; thread_idx++) {
       auto* stub = channels_[thread_idx % channels_.size()].get_stub();
       stream_[thread_idx] = stub->StreamingCall(&context_[thread_idx]);
       messages_issued_[thread_idx] = 0;
     }
-    StartThreads(num_threads_);
   }
-  ~SynchronousStreamingClient() {
+  ~SynchronousStreamingPingPongClient() {
     std::vector<std::thread> cleanup_threads;
     for (size_t i = 0; i < num_threads_; i++) {
       cleanup_threads.emplace_back([this, i]() {
         auto stream = &stream_[i];
         if (*stream) {
           (*stream)->WritesDone();
-          Status s = (*stream)->Finish();
-          if (!s.ok()) {
-            gpr_log(GPR_ERROR, "Stream %" PRIuPTR " received an error %s", i,
-                    s.error_message().c_str());
-          }
         }
       });
     }
-    for (size_t i = 0; i < num_threads_; i++) {
-      cleanup_threads[i].join();
+    for (auto& th : cleanup_threads) {
+      th.join();
     }
   }
 
@@ -176,7 +216,7 @@ class SynchronousStreamingClient final : public SynchronousClient {
     if (!WaitToIssue(thread_idx)) {
       return true;
     }
-    GPR_TIMER_SCOPE("SynchronousStreamingClient::ThreadFunc", 0);
+    GPR_TIMER_SCOPE("SynchronousStreamingPingPongClient::ThreadFunc", 0);
     double start = UsageTimer::Now();
     if (stream_[thread_idx]->Write(request_) &&
         stream_[thread_idx]->Read(&responses_[thread_idx])) {
@@ -192,40 +232,148 @@ class SynchronousStreamingClient final : public SynchronousClient {
       }
     }
     stream_[thread_idx]->WritesDone();
-    Status s = stream_[thread_idx]->Finish();
-    // don't set the value since this is either a failure (shouldn't be timed)
-    // or a stream-end (already has been timed)
-    entry->set_status(s.error_code());
-    if (!s.ok()) {
-      gpr_log(GPR_ERROR, "Stream %" PRIuPTR " received an error %s", thread_idx,
-              s.error_message().c_str());
-    }
+    FinishStream(entry, thread_idx);
     auto* stub = channels_[thread_idx % channels_.size()].get_stub();
-    context_[thread_idx].~ClientContext();
-    new (&context_[thread_idx]) ClientContext();
     stream_[thread_idx] = stub->StreamingCall(&context_[thread_idx]);
     messages_issued_[thread_idx] = 0;
     return true;
   }
+};
+
+class SynchronousStreamingFromClientClient final
+    : public SynchronousStreamingClient<grpc::ClientWriter<SimpleRequest>> {
+ public:
+  SynchronousStreamingFromClientClient(const ClientConfig& config)
+      : SynchronousStreamingClient(config), last_issue_(num_threads_) {
+    for (size_t thread_idx = 0; thread_idx < num_threads_; thread_idx++) {
+      auto* stub = channels_[thread_idx % channels_.size()].get_stub();
+      stream_[thread_idx] = stub->StreamingFromClient(&context_[thread_idx],
+                                                      &responses_[thread_idx]);
+      last_issue_[thread_idx] = UsageTimer::Now();
+    }
+  }
+  ~SynchronousStreamingFromClientClient() {
+    std::vector<std::thread> cleanup_threads;
+    for (size_t i = 0; i < num_threads_; i++) {
+      cleanup_threads.emplace_back([this, i]() {
+        auto stream = &stream_[i];
+        if (*stream) {
+          (*stream)->WritesDone();
+        }
+      });
+    }
+    for (auto& th : cleanup_threads) {
+      th.join();
+    }
+  }
+
+  bool ThreadFunc(HistogramEntry* entry, size_t thread_idx) override {
+    // Figure out how to make histogram sensible if this is rate-paced
+    if (!WaitToIssue(thread_idx)) {
+      return true;
+    }
+    GPR_TIMER_SCOPE("SynchronousStreamingFromClientClient::ThreadFunc", 0);
+    if (stream_[thread_idx]->Write(request_)) {
+      double now = UsageTimer::Now();
+      entry->set_value((now - last_issue_[thread_idx]) * 1e9);
+      last_issue_[thread_idx] = now;
+      return true;
+    }
+    stream_[thread_idx]->WritesDone();
+    FinishStream(entry, thread_idx);
+    auto* stub = channels_[thread_idx % channels_.size()].get_stub();
+    stream_[thread_idx] = stub->StreamingFromClient(&context_[thread_idx],
+                                                    &responses_[thread_idx]);
+    return true;
+  }
 
  private:
-  // These are both conceptually std::vector but cannot be for old compilers
-  // that expect contained classes to support copy constructors
-  std::vector<grpc::ClientContext> context_;
-  std::vector<
-      std::unique_ptr<grpc::ClientReaderWriter<SimpleRequest, SimpleResponse>>>
-      stream_;
-  const int messages_per_stream_;
-  std::vector<int> messages_issued_;
+  std::vector<double> last_issue_;
 };
 
-std::unique_ptr<Client> CreateSynchronousUnaryClient(
-    const ClientConfig& config) {
-  return std::unique_ptr<Client>(new SynchronousUnaryClient(config));
-}
-std::unique_ptr<Client> CreateSynchronousStreamingClient(
-    const ClientConfig& config) {
-  return std::unique_ptr<Client>(new SynchronousStreamingClient(config));
+class SynchronousStreamingFromServerClient final
+    : public SynchronousStreamingClient<grpc::ClientReader<SimpleResponse>> {
+ public:
+  SynchronousStreamingFromServerClient(const ClientConfig& config)
+      : SynchronousStreamingClient(config), last_recv_(num_threads_) {
+    for (size_t thread_idx = 0; thread_idx < num_threads_; thread_idx++) {
+      auto* stub = channels_[thread_idx % channels_.size()].get_stub();
+      stream_[thread_idx] =
+          stub->StreamingFromServer(&context_[thread_idx], request_);
+      last_recv_[thread_idx] = UsageTimer::Now();
+    }
+  }
+  bool ThreadFunc(HistogramEntry* entry, size_t thread_idx) override {
+    GPR_TIMER_SCOPE("SynchronousStreamingFromServerClient::ThreadFunc", 0);
+    if (stream_[thread_idx]->Read(&responses_[thread_idx])) {
+      double now = UsageTimer::Now();
+      entry->set_value((now - last_recv_[thread_idx]) * 1e9);
+      last_recv_[thread_idx] = now;
+      return true;
+    }
+    FinishStream(entry, thread_idx);
+    auto* stub = channels_[thread_idx % channels_.size()].get_stub();
+    stream_[thread_idx] =
+        stub->StreamingFromServer(&context_[thread_idx], request_);
+    return true;
+  }
+
+ private:
+  std::vector<double> last_recv_;
+};
+
+class SynchronousStreamingBothWaysClient final
+    : public SynchronousStreamingClient<
+          grpc::ClientReaderWriter<SimpleRequest, SimpleResponse>> {
+ public:
+  SynchronousStreamingBothWaysClient(const ClientConfig& config)
+      : SynchronousStreamingClient(config) {
+    for (size_t thread_idx = 0; thread_idx < num_threads_; thread_idx++) {
+      auto* stub = channels_[thread_idx % channels_.size()].get_stub();
+      stream_[thread_idx] = stub->StreamingBothWays(&context_[thread_idx]);
+    }
+  }
+  ~SynchronousStreamingBothWaysClient() {
+    std::vector<std::thread> cleanup_threads;
+    for (size_t i = 0; i < num_threads_; i++) {
+      cleanup_threads.emplace_back([this, i]() {
+        auto stream = &stream_[i];
+        if (*stream) {
+          (*stream)->WritesDone();
+        }
+      });
+    }
+    for (auto& th : cleanup_threads) {
+      th.join();
+    }
+  }
+
+  bool ThreadFunc(HistogramEntry* entry, size_t thread_idx) override {
+    // TODO (vjpai): Do this
+    return true;
+  }
+};
+
+std::unique_ptr<Client> CreateSynchronousClient(const ClientConfig& config) {
+  switch (config.rpc_type()) {
+    case UNARY:
+      return std::unique_ptr<Client>(new SynchronousUnaryClient(config));
+    case STREAMING:
+      return std::unique_ptr<Client>(
+          new SynchronousStreamingPingPongClient(config));
+    case STREAMING_FROM_CLIENT:
+      return std::unique_ptr<Client>(
+          new SynchronousStreamingFromClientClient(config));
+    case STREAMING_FROM_SERVER:
+      return std::unique_ptr<Client>(
+          new SynchronousStreamingFromServerClient(config));
+    case STREAMING_BOTH_WAYS:
+      return std::unique_ptr<Client>(
+          new SynchronousStreamingBothWaysClient(config));
+    default:
+      assert(false);
+      return nullptr;
+  }
 }
 
 }  // namespace testing

test/cpp/qps/qps_worker.cc

@@ -68,15 +68,11 @@ static std::unique_ptr<Client> CreateClient(const ClientConfig& config) {
 
   switch (config.client_type()) {
     case ClientType::SYNC_CLIENT:
-      return (config.rpc_type() == RpcType::UNARY)
-                 ? CreateSynchronousUnaryClient(config)
-                 : CreateSynchronousStreamingClient(config);
+      return CreateSynchronousClient(config);
     case ClientType::ASYNC_CLIENT:
-      return (config.rpc_type() == RpcType::UNARY)
-                 ? CreateAsyncUnaryClient(config)
-                 : (config.payload_config().has_bytebuf_params()
-                        ? CreateGenericAsyncStreamingClient(config)
-                        : CreateAsyncStreamingClient(config));
+      return config.payload_config().has_bytebuf_params()
+                 ? CreateGenericAsyncStreamingClient(config)
+                 : CreateAsyncClient(config);
     default:
       abort();
   }

test/cpp/qps/server_async.cc

@@ -71,6 +71,18 @@ class AsyncQpsServerTest final : public grpc::testing::Server {
                          ServerAsyncReaderWriter<ResponseType, RequestType> *,
                          CompletionQueue *, ServerCompletionQueue *, void *)>
           request_streaming_function,
+      std::function<void(ServiceType *, ServerContextType *,
+                         ServerAsyncReader<ResponseType, RequestType> *,
+                         CompletionQueue *, ServerCompletionQueue *, void *)>
+          request_streaming_from_client_function,
+      std::function<void(ServiceType *, ServerContextType *, RequestType *,
+                         ServerAsyncWriter<ResponseType> *, CompletionQueue *,
+                         ServerCompletionQueue *, void *)>
+          request_streaming_from_server_function,
+      std::function<void(ServiceType *, ServerContextType *,
+                         ServerAsyncReaderWriter<ResponseType, RequestType> *,
+                         CompletionQueue *, ServerCompletionQueue *, void *)>
+          request_streaming_both_ways_function,
       std::function<grpc::Status(const PayloadConfig &, const RequestType *,
                                  ResponseType *)>
           process_rpc)
@@ -107,7 +119,7 @@ class AsyncQpsServerTest final : public grpc::testing::Server {
         std::bind(process_rpc, config.payload_config(), std::placeholders::_1,
                   std::placeholders::_2);
 
-    for (int i = 0; i < 15000; i++) {
+    for (int i = 0; i < 5000; i++) {
       for (int j = 0; j < num_threads; j++) {
         if (request_unary_function) {
           auto request_unary = std::bind(
@@ -125,6 +137,26 @@ class AsyncQpsServerTest final : public grpc::testing::Server {
           contexts_.emplace_back(new ServerRpcContextStreamingImpl(
               request_streaming, process_rpc_bound));
         }
+        if (request_streaming_from_client_function) {
+          auto request_streaming_from_client = std::bind(
+              request_streaming_from_client_function, &async_service_,
+              std::placeholders::_1, std::placeholders::_2, srv_cqs_[j].get(),
+              srv_cqs_[j].get(), std::placeholders::_3);
+          contexts_.emplace_back(new ServerRpcContextStreamingFromClientImpl(
+              request_streaming_from_client, process_rpc_bound));
+        }
+        if (request_streaming_from_server_function) {
+          auto request_streaming_from_server =
+              std::bind(request_streaming_from_server_function, &async_service_,
+                        std::placeholders::_1, std::placeholders::_2,
+                        std::placeholders::_3, srv_cqs_[j].get(),
+                        srv_cqs_[j].get(), std::placeholders::_4);
+          contexts_.emplace_back(new ServerRpcContextStreamingFromServerImpl(
+              request_streaming_from_server, process_rpc_bound));
+        }
+        if (request_streaming_both_ways_function) {
+          // TODO(vjpai): Add this code
+        }
       }
     }
 
@@ -289,8 +321,8 @@ class AsyncQpsServerTest final : public grpc::testing::Server {
       if (!ok) {
         return false;
       }
-      stream_.Read(&req_, AsyncQpsServerTest::tag(this));
       next_state_ = &ServerRpcContextStreamingImpl::read_done;
+      stream_.Read(&req_, AsyncQpsServerTest::tag(this));
       return true;
     }
 
@@ -300,23 +332,23 @@ class AsyncQpsServerTest final : public grpc::testing::Server {
         // Call the RPC processing function
         grpc::Status status = invoke_method_(&req_, &response_);
         // initiate the write
-        stream_.Write(response_, AsyncQpsServerTest::tag(this));
         next_state_ = &ServerRpcContextStreamingImpl::write_done;
+        stream_.Write(response_, AsyncQpsServerTest::tag(this));
       } else {  // client has sent writes done
         // finish the stream
-        stream_.Finish(Status::OK, AsyncQpsServerTest::tag(this));
         next_state_ = &ServerRpcContextStreamingImpl::finish_done;
+        stream_.Finish(Status::OK, AsyncQpsServerTest::tag(this));
       }
       return true;
     }
     bool write_done(bool ok) {
       // now go back and get another streaming read!
       if (ok) {
-        stream_.Read(&req_, AsyncQpsServerTest::tag(this));
         next_state_ = &ServerRpcContextStreamingImpl::read_done;
+        stream_.Read(&req_, AsyncQpsServerTest::tag(this));
       } else {
-        stream_.Finish(Status::OK, AsyncQpsServerTest::tag(this));
         next_state_ = &ServerRpcContextStreamingImpl::finish_done;
+        stream_.Finish(Status::OK, AsyncQpsServerTest::tag(this));
       }
       return true;
     }
@@ -335,6 +367,146 @@ class AsyncQpsServerTest final : public grpc::testing::Server {
     grpc::ServerAsyncReaderWriter<ResponseType, RequestType> stream_;
   };
 
+  class ServerRpcContextStreamingFromClientImpl final
+      : public ServerRpcContext {
+   public:
+    ServerRpcContextStreamingFromClientImpl(
+        std::function<void(ServerContextType *,
+                           grpc::ServerAsyncReader<ResponseType, RequestType> *,
+                           void *)>
+            request_method,
+        std::function<grpc::Status(const RequestType *, ResponseType *)>
+            invoke_method)
+        : srv_ctx_(new ServerContextType),
+          next_state_(&ServerRpcContextStreamingFromClientImpl::request_done),
+          request_method_(request_method),
+          invoke_method_(invoke_method),
+          stream_(srv_ctx_.get()) {
+      request_method_(srv_ctx_.get(), &stream_, AsyncQpsServerTest::tag(this));
+    }
+    ~ServerRpcContextStreamingFromClientImpl() override {}
+    bool RunNextState(bool ok) override { return (this->*next_state_)(ok); }
+    void Reset() override {
+      srv_ctx_.reset(new ServerContextType);
+      req_ = RequestType();
+      stream_ =
+          grpc::ServerAsyncReader<ResponseType, RequestType>(srv_ctx_.get());
+
+      // Then request the method
+      next_state_ = &ServerRpcContextStreamingFromClientImpl::request_done;
+      request_method_(srv_ctx_.get(), &stream_, AsyncQpsServerTest::tag(this));
+    }
+
+   private:
+    bool request_done(bool ok) {
+      if (!ok) {
+        return false;
+      }
+      next_state_ = &ServerRpcContextStreamingFromClientImpl::read_done;
+      stream_.Read(&req_, AsyncQpsServerTest::tag(this));
+      return true;
+    }
+
+    bool read_done(bool ok) {
+      if (ok) {
+        // In this case, just do another read
+        // next_state_ is unchanged
+        stream_.Read(&req_, AsyncQpsServerTest::tag(this));
+        return true;
+      } else {  // client has sent writes done
+        // invoke the method
+        // Call the RPC processing function
+        grpc::Status status = invoke_method_(&req_, &response_);
+        // finish the stream
+        next_state_ = &ServerRpcContextStreamingFromClientImpl::finish_done;
+        stream_.Finish(response_, Status::OK, AsyncQpsServerTest::tag(this));
+      }
+      return true;
+    }
+    bool finish_done(bool ok) { return false; /* reset the context */ }
+
+    std::unique_ptr<ServerContextType> srv_ctx_;
+    RequestType req_;
+    ResponseType response_;
+    bool (ServerRpcContextStreamingFromClientImpl::*next_state_)(bool);
+    std::function<void(ServerContextType *,
+                       grpc::ServerAsyncReader<ResponseType, RequestType> *,
+                       void *)>
+        request_method_;
+    std::function<grpc::Status(const RequestType *, ResponseType *)>
+        invoke_method_;
+    grpc::ServerAsyncReader<ResponseType, RequestType> stream_;
+  };
+
+  class ServerRpcContextStreamingFromServerImpl final
+      : public ServerRpcContext {
+   public:
+    ServerRpcContextStreamingFromServerImpl(
+        std::function<void(ServerContextType *, RequestType *,
+                           grpc::ServerAsyncWriter<ResponseType> *, void *)>
+            request_method,
+        std::function<grpc::Status(const RequestType *, ResponseType *)>
+            invoke_method)
+        : srv_ctx_(new ServerContextType),
+          next_state_(&ServerRpcContextStreamingFromServerImpl::request_done),
+          request_method_(request_method),
+          invoke_method_(invoke_method),
+          stream_(srv_ctx_.get()) {
+      request_method_(srv_ctx_.get(), &req_, &stream_,
+                      AsyncQpsServerTest::tag(this));
+    }
+    ~ServerRpcContextStreamingFromServerImpl() override {}
+    bool RunNextState(bool ok) override { return (this->*next_state_)(ok); }
+    void Reset() override {
+      srv_ctx_.reset(new ServerContextType);
+      req_ = RequestType();
+      stream_ = grpc::ServerAsyncWriter<ResponseType>(srv_ctx_.get());
+
+      // Then request the method
+      next_state_ = &ServerRpcContextStreamingFromServerImpl::request_done;
+      request_method_(srv_ctx_.get(), &req_, &stream_,
+                      AsyncQpsServerTest::tag(this));
+    }
+
+   private:
+    bool request_done(bool ok) {
+      if (!ok) {
+        return false;
+      }
+      // invoke the method
+      // Call the RPC processing function
+      grpc::Status status = invoke_method_(&req_, &response_);
+
+      next_state_ = &ServerRpcContextStreamingFromServerImpl::write_done;
+      stream_.Write(response_, AsyncQpsServerTest::tag(this));
+      return true;
+    }
+
+    bool write_done(bool ok) {
+      if (ok) {
+        // Do another write!
+        // next_state_ is unchanged
+        stream_.Write(response_, AsyncQpsServerTest::tag(this));
+      } else {  // must be done so let's finish
+        next_state_ = &ServerRpcContextStreamingFromServerImpl::finish_done;
+        stream_.Finish(Status::OK, AsyncQpsServerTest::tag(this));
+      }
+      return true;
+    }
+    bool finish_done(bool ok) { return false; /* reset the context */ }
+
+    std::unique_ptr<ServerContextType> srv_ctx_;
+    RequestType req_;
+    ResponseType response_;
+    bool (ServerRpcContextStreamingFromServerImpl::*next_state_)(bool);
+    std::function<void(ServerContextType *, RequestType *,
+                       grpc::ServerAsyncWriter<ResponseType> *, void *)>
+        request_method_;
+    std::function<grpc::Status(const RequestType *, ResponseType *)>
+        invoke_method_;
+    grpc::ServerAsyncWriter<ResponseType> stream_;
+  };
+
   std::vector<std::thread> threads_;
   std::unique_ptr<grpc::Server> server_;
   std::vector<std::unique_ptr<grpc::ServerCompletionQueue>> srv_cqs_;
@@ -390,6 +562,9 @@ std::unique_ptr<Server> CreateAsyncServer(const ServerConfig &config) {
           config, RegisterBenchmarkService,
           &BenchmarkService::AsyncService::RequestUnaryCall,
           &BenchmarkService::AsyncService::RequestStreamingCall,
+          &BenchmarkService::AsyncService::RequestStreamingFromClient,
+          &BenchmarkService::AsyncService::RequestStreamingFromServer,
+          &BenchmarkService::AsyncService::RequestStreamingBothWays,
           ProcessSimpleRPC));
 }
 std::unique_ptr<Server> CreateAsyncGenericServer(const ServerConfig &config) {
@@ -397,7 +572,8 @@ std::unique_ptr<Server> CreateAsyncGenericServer(const ServerConfig &config) {
       new AsyncQpsServerTest<ByteBuffer, ByteBuffer, grpc::AsyncGenericService,
                              grpc::GenericServerContext>(
           config, RegisterGenericService, nullptr,
-          &grpc::AsyncGenericService::RequestCall, ProcessGenericRPC));
+          &grpc::AsyncGenericService::RequestCall, nullptr, nullptr, nullptr,
+          ProcessGenericRPC));
 }
 
 }  // namespace testing

test/cpp/qps/server_sync.cc

@@ -31,6 +31,9 @@
  *
  */
 
+#include <atomic>
+#include <thread>
+
 #include <grpc++/resource_quota.h>
 #include <grpc++/security/server_credentials.h>
 #include <grpc++/server.h>
@@ -52,12 +55,9 @@ class BenchmarkServiceImpl final : public BenchmarkService::Service {
  public:
   Status UnaryCall(ServerContext* context, const SimpleRequest* request,
                    SimpleResponse* response) override {
-    if (request->response_size() > 0) {
-      if (!Server::SetPayload(request->response_type(),
-                              request->response_size(),
-                              response->mutable_payload())) {
-        return Status(grpc::StatusCode::INTERNAL, "Error creating payload.");
-      }
+    auto s = SetResponse(request, response);
+    if (!s.ok()) {
+      return s;
     }
     return Status::OK;
   }
@@ -67,12 +67,9 @@ class BenchmarkServiceImpl final : public BenchmarkService::Service {
     SimpleRequest request;
     while (stream->Read(&request)) {
       SimpleResponse response;
-      if (request.response_size() > 0) {
-        if (!Server::SetPayload(request.response_type(),
-                                request.response_size(),
-                                response.mutable_payload())) {
-          return Status(grpc::StatusCode::INTERNAL, "Error creating payload.");
-        }
+      auto s = SetResponse(&request, &response);
+      if (!s.ok()) {
+        return s;
       }
       if (!stream->Write(response)) {
         return Status(StatusCode::INTERNAL, "Server couldn't respond");
@@ -80,6 +77,96 @@ class BenchmarkServiceImpl final : public BenchmarkService::Service {
     }
     return Status::OK;
   }
+  Status StreamingFromClient(ServerContext* context,
+                             ServerReader<SimpleRequest>* stream,
+                             SimpleResponse* response) override {
+    auto s = ClientPull(context, stream, response);
+    if (!s.ok()) {
+      return s;
+    }
+    return Status::OK;
+  }
+  Status StreamingFromServer(ServerContext* context,
+                             const SimpleRequest* request,
+                             ServerWriter<SimpleResponse>* stream) override {
+    SimpleResponse response;
+    auto s = SetResponse(request, &response);
+    if (!s.ok()) {
+      return s;
+    }
+    return ServerPush(context, stream, response, nullptr);
+  }
+  Status StreamingBothWays(
+      ServerContext* context,
+      ServerReaderWriter<SimpleResponse, SimpleRequest>* stream) override {
+    // Read the first client message to setup server response
+    SimpleRequest request;
+    if (!stream->Read(&request)) {
+      return Status::OK;
+    }
+    SimpleResponse response;
+    auto s = SetResponse(&request, &response);
+    if (!s.ok()) {
+      return s;
+    }
+    std::atomic_bool done;
+    Status sp;
+    std::thread t([context, stream, &response, &done, &sp]() {
+      sp = ServerPush(context, stream, response, [&done]() {
+        return done.load(std::memory_order_relaxed);
+      });
+    });
+    SimpleResponse dummy;
+    auto cp = ClientPull(context, stream, &dummy);
+    done.store(true, std::memory_order_relaxed);  // can be lazy
+    t.join();
+    if (!cp.ok()) {
+      return cp;
+    }
+    if (!sp.ok()) {
+      return sp;
+    }
+    return Status::OK;
+  }
+
+ private:
+  static Status ClientPull(ServerContext* context,
+                           ReaderInterface<SimpleRequest>* stream,
+                           SimpleResponse* response) {
+    SimpleRequest request;
+    while (stream->Read(&request)) {
+    }
+    if (request.response_size() > 0) {
+      if (!Server::SetPayload(request.response_type(), request.response_size(),
+                              response->mutable_payload())) {
+        return Status(grpc::StatusCode::INTERNAL, "Error creating payload.");
+      }
+    }
+    return Status::OK;
+  }
+  static Status ServerPush(ServerContext* context,
+                           WriterInterface<SimpleResponse>* stream,
+                           const SimpleResponse& response,
+                           std::function<bool()> done) {
+    while ((done == nullptr) || !done()) {
+      // TODO(vjpai): Add potential for rate-pacing on this
+      if (!stream->Write(response)) {
+        return Status(StatusCode::INTERNAL, "Server couldn't push");
+      }
+    }
+    return Status::OK;
+  }
+  static Status SetResponse(const SimpleRequest* request,
+                            SimpleResponse* response) {
+    if (request->response_size() > 0) {
+      if (!Server::SetPayload(request->response_type(),
+                              request->response_size(),
+                              response->mutable_payload())) {
+        return Status(grpc::StatusCode::INTERNAL, "Error creating payload.");
+      }
+    }
+    return Status::OK;
+  }
 };
 
 class SynchronousServer final : public grpc::testing::Server {

tools/profiling/microbenchmarks/bm_diff.py

@@ -56,6 +56,10 @@ _INTERESTING = (
   'writes_per_iteration',
   'atm_cas_per_iteration',
   'atm_add_per_iteration',
+  'cli_transport_stalls_per_iteration',
+  'cli_stream_stalls_per_iteration',
+  'svr_transport_stalls_per_iteration',
+  'svr_stream_stalls_per_iteration'
   'nows_per_iteration',
 )
 

tools/run_tests/performance/scenario_config.py

@@ -310,7 +310,7 @@ class CXXLanguage:
         secure=secure,
         categories=smoketest_categories + [SCALABLE])
 
-      for rpc_type in ['unary', 'streaming']:
+      for rpc_type in ['unary', 'streaming', 'streaming_from_client', 'streaming_from_server']:
         for synchronicity in ['sync', 'async']:
           yield _ping_pong_scenario(
               'cpp_protobuf_%s_%s_ping_pong_%s' % (synchronicity, rpc_type, secstr),

tools/run_tests/python_utils/port_server.py

@@ -1,4 +1,4 @@
-#!/usr/bin/env python
+#!/usr/bin/env python2.7
 # Copyright 2015, Google Inc.
 # All rights reserved.
 #
@@ -30,8 +30,6 @@
 
 """Manage TCP ports for unit tests; started by run_tests.py"""
 
-from __future__ import print_function
-
 import argparse
 from BaseHTTPServer import HTTPServer, BaseHTTPRequestHandler
 import hashlib
@@ -42,16 +40,17 @@ import time
 import random
 from SocketServer import ThreadingMixIn
 import threading
+import platform
 
 
 # increment this number whenever making a change to ensure that
 # the changes are picked up by running CI servers
 # note that all changes must be backwards compatible
-_MY_VERSION = 19
+_MY_VERSION = 20
 
 
 if len(sys.argv) == 2 and sys.argv[1] == 'dump_version':
-  print(_MY_VERSION)
+  print _MY_VERSION
   sys.exit(0)
 
 
@@ -67,13 +66,16 @@ if args.logfile is not None:
   sys.stderr = open(args.logfile, 'w')
   sys.stdout = sys.stderr
 
-print('port server running on port %d' % args.port)
+print 'port server running on port %d' % args.port
 
 pool = []
 in_use = {}
 mu = threading.Lock()
 
 def can_connect(port):
+  # this test is only really useful on unices where SO_REUSE_PORT is available
+  # so on Windows, where this test is expensive, skip it
+  if platform.system() == 'Windows': return False
   s = socket.socket()
   try:
     s.connect(('localhost', port))
@@ -137,7 +139,7 @@ keep_running = True
 
 
 class Handler(BaseHTTPRequestHandler):
-  
+
   def setup(self):
     # If the client is unreachable for 5 seconds, close the connection
     self.timeout = 5
@@ -195,6 +197,4 @@ class Handler(BaseHTTPRequestHandler):
 class ThreadedHTTPServer(ThreadingMixIn, HTTPServer):
   """Handle requests in a separate thread"""
 
-
 ThreadedHTTPServer(('', args.port), Handler).serve_forever()
-

tools/run_tests/python_utils/start_port_server.py

@@ -27,9 +27,7 @@
 # (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 # OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
-from __future__ import print_function
-
-from six.moves import urllib
+import urllib
 import jobset
 import logging
 import os
@@ -50,9 +48,9 @@ def start_port_server():
     # otherwise, leave it up
     try:
         version = int(
-            urllib.request.urlopen(
-                'http://localhost:%d/version_number' % _PORT_SERVER_PORT,
-                timeout=10).read())
+            urllib.urlopen(
+                'http://localhost:%d/version_number' %
+                _PORT_SERVER_PORT).read())
         logging.info('detected port server running version %d', version)
         running = True
     except Exception as e:
@@ -69,8 +67,8 @@ def start_port_server():
         running = (version >= current_version)
         if not running:
             logging.info('port_server version mismatch: killing the old one')
-            urllib.request.urlopen('http://localhost:%d/quitquitquit' %
-                                   _PORT_SERVER_PORT).read()
+            urllib.urlopen('http://localhost:%d/quitquitquit' %
+                           _PORT_SERVER_PORT).read()
             time.sleep(1)
     if not running:
         fd, logfile = tempfile.mkstemp()
@@ -109,9 +107,8 @@ def start_port_server():
                 # try one final time: maybe another build managed to start one
                 time.sleep(1)
                 try:
-                    urllib.request.urlopen(
-                        'http://localhost:%d/get' % _PORT_SERVER_PORT,
-                        timeout=1).read()
+                    urllib.urlopen(
+                        'http://localhost:%d/get' % _PORT_SERVER_PORT).read()
                     logging.info(
                         'last ditch attempt to contact port server succeeded')
                     break
@@ -119,18 +116,18 @@ def start_port_server():
                     logging.exception(
                         'final attempt to contact port server failed')
                     port_log = open(logfile, 'r').read()
-                    print(port_log)
+                    print port_log
                     sys.exit(1)
             try:
                 port_server_url = 'http://localhost:%d/get' % _PORT_SERVER_PORT
-                urllib.request.urlopen(port_server_url, timeout=1).read()
+                urllib.urlopen(port_server_url).read()
                 logging.info('port server is up and ready')
                 break
             except socket.timeout:
                 logging.exception('while waiting for port_server')
                 time.sleep(1)
                 waits += 1
-            except urllib.error.URLError:
+            except IOError:
                 logging.exception('while waiting for port_server')
                 time.sleep(1)
                 waits += 1

tools/run_tests/start_port_server.py

@@ -1,4 +1,4 @@
-#!/usr/bin/env python
+#!/usr/bin/env python2.7
 
 # Copyright 2017, Google Inc.
 # All rights reserved.
@@ -39,10 +39,8 @@ The path to this file is called out in test/core/util/port.c, and printed as
 an error message to users.
 """
 
-from __future__ import print_function
-
 import python_utils.start_port_server as start_port_server
 
 start_port_server.start_port_server()
 
-print("Port server started successfully")
+print "Port server started successfully"