3a519a0b64
Replaced grpc_core New & Delete with C++ new & delete
5 years ago
406b70629a
Remove unused parameter warning (17 of 20)
5 years ago
0b06676c9e
hpack encoder optimizations.
...
Removed some cycles and branches from hpack_enc for CH2.
Specifically:
1. Pushed certain metadata key/value length checks to
prepare_application_metadata() in src/core/lib/surface/call.cc.
This means that rather than check all key/val lengths for all metadata, we only
do so for custom added user metadata. Inside CH2, we change the length checks to
debug checks so we can catch if core/filter metadata fails to pass the check.
2. Changed various asserts to debug asserts when able.
3. Refactored some of the header emission code to remove duplicated code.
4. Un-inlined some logging methods.
This results in somewhat faster hpack_encoder performance:
BM_HpackEncoderInitDestroy
222ns ± 0% 221ns ± 0% -0.29% (p=0.000 n=34+34)
BM_HpackEncoderEncodeDeadline
[framing_bytes/iter:9 header_bytes/iter:6 ] 135ns ± 1%
124ns ± 0% -8.05% (p=0.000 n=39+38)
BM_HpackEncoderEncodeHeader<EmptyBatch>/0/16384
[framing_bytes/iter:9 header_bytes/iter:0 ] 34.2ns ± 0%
34.2ns ± 0% -0.01% (p=0.014 n=34+38)
BM_HpackEncoderEncodeHeader<EmptyBatch>/1/16384
[framing_bytes/iter:9 header_bytes/iter:0 ] 34.2ns ± 0%
34.2ns ± 0% -0.04% (p=0.004 n=34+37)
BM_HpackEncoderEncodeHeader<SingleStaticElem>/0/16384
[framing_bytes/iter:9 header_bytes/iter:1 ] 47.5ns ± 0%
45.9ns ± 0% -3.28% (p=0.000 n=28+38)
BM_HpackEncoderEncodeHeader<SingleInternedKeyElem>/0/16384
[framing_bytes/iter:9 header_bytes/iter:6 ] 77.0ns ± 1%
68.3ns ± 1% -11.33% (p=0.000 n=39+40)
BM_HpackEncoderEncodeHeader<SingleInternedElem>/0/16384
[framing_bytes/iter:9 header_bytes/iter:1 ] 47.7ns ± 1%
45.5ns ± 0% -4.63% (p=0.000 n=39+33)
BM_HpackEncoderEncodeHeader<SingleInternedBinaryElem<1, false>>/0/16384
[framing_bytes/iter:9 header_bytes/iter:1 ] 47.2ns ± 0%
45.3ns ± 0% -3.96% (p=0.000 n=33+34)
BM_HpackEncoderEncodeHeader<SingleInternedBinaryElem<3, false>>/0/16384
[framing_bytes/iter:9 header_bytes/iter:1 ] 47.7ns ± 0%
45.6ns ± 0% -4.54% (p=0.000 n=38+40)
BM_HpackEncoderEncodeHeader<SingleInternedBinaryElem<10, false>>/0/16384
[framing_bytes/iter:9 header_bytes/iter:1 ] 47.7ns ± 0%
45.5ns ± 0% -4.63% (p=0.000 n=39+32)
BM_HpackEncoderEncodeHeader<SingleInternedBinaryElem<31, false>>/0/16384
[framing_bytes/iter:9 header_bytes/iter:1 ] 47.8ns ± 0%
45.6ns ± 1% -4.59% (p=0.000 n=38+39)
BM_HpackEncoderEncodeHeader<SingleInternedBinaryElem<100, false>>/0/16384
[framing_bytes/iter:9 header_bytes/iter:1 ] 47.8ns ± 0%
45.5ns ± 0% -4.64% (p=0.000 n=39+36)
BM_HpackEncoderEncodeHeader<SingleInternedBinaryElem<1, true>>/0/16384
[framing_bytes/iter:9 header_bytes/iter:1 ] 47.3ns ± 0%
45.3ns ± 0% -4.09% (p=0.000 n=38+36)
BM_HpackEncoderEncodeHeader<SingleInternedBinaryElem<3, true>>/0/16384
[framing_bytes/iter:9 header_bytes/iter:1 ] 47.8ns ± 1%
45.6ns ± 0% -4.71% (p=0.000 n=37+40)
BM_HpackEncoderEncodeHeader<SingleInternedBinaryElem<10, true>>/0/16384
[framing_bytes/iter:9 header_bytes/iter:1 ] 47.7ns ± 0%
45.5ns ± 0% -4.66% (p=0.000 n=39+32)
BM_HpackEncoderEncodeHeader<SingleInternedBinaryElem<31, true>>/0/16384
[framing_bytes/iter:9 header_bytes/iter:1 ] 47.8ns ± 1%
45.6ns ± 1% -4.62% (p=0.000 n=37+39)
BM_HpackEncoderEncodeHeader<SingleInternedBinaryElem<100, true>>/0/16384
[framing_bytes/iter:9 header_bytes/iter:1 ] 47.7ns ± 0%
45.5ns ± 0% -4.67% (p=0.000 n=38+32)
BM_HpackEncoderEncodeHeader<SingleNonInternedElem>/0/16384
[framing_bytes/iter:9 header_bytes/iter:9 ] 80.5ns ± 1%
74.7ns ± 0% -7.16% (p=0.000 n=38+35)
BM_HpackEncoderEncodeHeader<SingleNonInternedBinaryElem<1, false>>/0/16384
[framing_bytes/iter:9 header_bytes/iter:12 ] 105ns ± 1%
99ns ± 0% -5.91% (p=0.000 n=38+34)
BM_HpackEncoderEncodeHeader<SingleNonInternedBinaryElem<3, false>>/0/16384
[framing_bytes/iter:9 header_bytes/iter:14 ] 111ns ± 1%
106ns ± 1% -4.86% (p=0.020 n=39+2)
BM_HpackEncoderEncodeHeader<SingleNonInternedBinaryElem<10, false>>/0/16384
[framing_bytes/iter:9 header_bytes/iter:23 ] 135ns ± 0%
130ns ± 0% -3.45% (p=0.020 n=35+2)
BM_HpackEncoderEncodeHeader<SingleNonInternedBinaryElem<31, false>>/0/16384
[framing_bytes/iter:9 header_bytes/iter:46 ] 225ns ± 1%
223ns ± 0% -0.91% (p=0.003 n=37+2)
BM_HpackEncoderEncodeHeader<SingleNonInternedBinaryElem<100, false>>/0/16384
[framing_bytes/iter:9 header_bytes/iter:120 ] 467ns ± 0%
472ns ± 0% +1.09% (p=0.003 n=38+2)
BM_HpackEncoderEncodeHeader<SingleNonInternedBinaryElem<1, true>>/0/16384
[framing_bytes/iter:9 header_bytes/iter:12 ] 81.6ns ± 1%
74.8ns ± 0% -8.40% (p=0.000 n=37+33)
BM_HpackEncoderEncodeHeader<SingleNonInternedBinaryElem<3, true>>/0/16384
[framing_bytes/iter:9 header_bytes/iter:14 ] 82.0ns ± 1%
74.8ns ± 0% -8.80% (p=0.000 n=37+32)
BM_HpackEncoderEncodeHeader<SingleNonInternedBinaryElem<10, true>>/0/16384
[framing_bytes/iter:9 header_bytes/iter:21 ] 82.1ns ± 1%
74.9ns ± 0% -8.86% (p=0.000 n=35+34)
BM_HpackEncoderEncodeHeader<SingleNonInternedBinaryElem<31, true>>/0/16384
[framing_bytes/iter:9 header_bytes/iter:42 ] 97.6ns ± 2%
91.8ns ± 0% -5.95% (p=0.000 n=35+27)
BM_HpackEncoderEncodeHeader<SingleNonInternedBinaryElem<100, true>>/0/16384
[framing_bytes/iter:9 header_bytes/iter:111 ] 97.2ns ± 1%
91.2ns ± 2% -6.19% (p=0.000 n=37+38)
BM_HpackEncoderEncodeHeader<SingleNonInternedElem>/0/1
[framing_bytes/iter:54 header_bytes/iter:9 ] 230ns ± 0%
221ns ± 0% -3.91% (p=0.000 n=38+37)
BM_HpackEncoderEncodeHeader<MoreRepresentativeClientInitialMetadata>/0/16384
[framing_bytes/iter:9 header_bytes/iter:16 ] 206ns ± 2%
170ns ± 1% -17.51% (p=0.000 n=39+39)
BM_HpackEncoderEncodeHeader<RepresentativeServerInitialMetadata>/0/16384
[framing_bytes/iter:9 header_bytes/iter:3 ] 66.4ns ± 2%
62.5ns ± 1% -5.85% (p=0.000 n=34+39)
BM_HpackEncoderEncodeHeader<RepresentativeServerTrailingMetadata>/1/16384
[framing_bytes/iter:9 header_bytes/iter:1 ] 47.5ns ± 0%
45.9ns ± 1% -3.29% (p=0.000 n=26+38)
5 years ago
1077b3435c
Use range-based for on state rather than state.KeepRunning when possible
5 years ago
b46e3668d3
s/branch/tail_call/ for CH2 on_hdr().
...
on_hdr() checks if a void-return function pointer is null before jumping to it.
If it is null, it returns an error; else it executes that function and returns
success.
This change converts the void-returning function to one that returns a
grpc_error* and thus saves a branch in on_hdr() (since we're branching once by
following the function pointer anyways, we're effectively coalescing these two
branches).
5 years ago
59564ebd96
Fix warnings to unblock gcc8 support
5 years ago
b1d73a01f1
Removed duplicate static table from hpack table. Removed an or instruction for
...
every usage of static grpc metadata. Inlined hpack table lookups for static
metadata.
This leads to faster hpack parser creation:
BM_HpackParserInitDestroy 5.32µs ± 1% 0.06µs ± 1% -98.91% (p=0.000 n=18+19)
And slightly faster parsing:
BM_HpackParserParseHeader<RepresentativeClientInitialMetadata, OnInitialHeader>
456ns ± 1% 435ns ± 1% -4.74% (p=0.000 n=18+19)
BM_HpackParserParseHeader<MoreRepresentativeClientInitialMetadata,
OnInitialHeader>
1.06µs ± 2% 1.04µs ± 2% -1.82% (p=0.000 n=19+20)
It also yields a slight (0.5 - 1.0 microsecond) reduction in CPU time for
fullstack unary pingpong:
BM_UnaryPingPong<TCP, NoOpMutator, NoOpMutator>/0/512
[polls/iter:3.0001 ] 23.9µs ± 2%
23.0µs ± 1% -3.63% (p=0.002 n=6+6)
BM_UnaryPingPong<TCP, NoOpMutator, NoOpMutator>/0/32768
[polls/iter:3.00015 ] 35.1µs ± 1%
34.2µs ± 1% -2.57% (p=0.036 n=5+3)
BM_UnaryPingPong<MinTCP, NoOpMutator, NoOpMutator>/8/0
[polls/iter:3.00011 ] 21.7µs ± 3%
21.2µs ± 2% -2.44% (p=0.017 n=6+5)
5 years ago
e18ed03c04
Made gRPC inialized after entering main function in microbenchmarks.
6 years ago
8ce42f67b2
Shrink arena size by 40 bytes and add additional
...
alignment options (for cache-alignment).
We shrink by:
1) Removing an unnecessary zone pointer.
2) Replacing gpr_mu (40 bytes when using pthread_mutex_t) with
std::atomic_flag.
We also header-inline the fastpath alloc (ie. when not doing a zone
alloc) and move the malloc() for a zone alloc outside of the mutex
critical zone, which allows us to replace the mutex with a spinlock.
We also cache-align created arenas.
6 years ago
48e4a81b05
Remeve memset(0) from arena allocated memory.
...
Callers are updated to properly initialize the memory.
This behavior can be overridden using GRPC_ARENA_INIT_STRATEGY
environment variable.
6 years ago
4c6e7ce15d
Destroy metadata buffer at end of benchmark loop
6 years ago
d44feec92f
Reassign arena pointer instead of stomping on memory
6 years ago
4b721fbde0
Destroy arena at end of benchmark to not leak memory
6 years ago
29d9489ea9
Increase initial arena size to be more representative of real workload scenario and increase frequency of recreating the arena to avoid oom
6 years ago
91727bd015
Move arena create outside of benchmark, format, and typo fix
6 years ago
967bbcd5d3
Fixing benchmark name and adding a new one
6 years ago
58e0cbf9fb
Enable the performance-* clang-tidy checks
7 years ago
40ec89ff67
Support microbenchmarks internally
7 years ago
2f4161c210
Use stack frame size limits for consistency with internal builds
7 years ago
4d20a66685
Run clang fmt
7 years ago
be82e64b3d
Autofix c casts to c++ casts
7 years ago
38fcd0c6c3
clang-format
7 years ago
8cf1470a51
Revert "Revert "All instances of exec_ctx being passed around in src/core removed""
7 years ago
ad4d2dde00
Revert "All instances of exec_ctx being passed around in src/core removed"
7 years ago
c354269ba7
Remove _ prefixed variable names
7 years ago
6c26b16fe0
Move ExecCtx to grpc_core namespace. Make exec_ctx a private static in ExecCtx and some minor changes
7 years ago
75122c2357
Address some PR comments
7 years ago
4ac2b8e585
Enable clang-tidy as a sanity check, fix up all known failures
7 years ago
0032548674
Correction to closure.cc,bm_chttp2_hpack and few more
7 years ago
3150744c71
Removing more exec_ctx instances
7 years ago
baa14a975e
Update clang-format to 5.0
7 years ago
39aed1ae8b
Remove unnecessary extern Cs
7 years ago
c296e82e11
clang fmt
7 years ago
9e3eedb6af
Remove old header benchmark
7 years ago
6bf4bcef04
Fix bm_diff
7 years ago
83085aa74f
Add a microbm, seeing 195ns with current impl and 162ns with new impl
7 years ago
0ee7574732
Removing instances of exec_ctx being passed around in functions in
...
src/core. exec_ctx is now a thread_local pointer of type ExecCtx instead of
grpc_exec_ctx which is initialized whenever ExecCtx is instantiated. ExecCtx
also keeps track of the previous exec_ctx so that nesting of exec_ctx is
allowed. This means that there is only one exec_ctx being used at any
time. Also, grpc_exec_ctx_finish is called in the destructor of the
object, and the previous exec_ctx is restored to avoid breaking current
functionality. The code still explicitly calls grpc_exec_ctx_finish
because removing all such instances causes the code to break.
7 years ago
c010d1d18a
Update benchmark according to new encoding method
7 years ago
c94c7cc5b5
restore existing fixtures
7 years ago
377636f4d2
Make hpack micro bm more representative
7 years ago
bd3b93b4b5
Add support for Trailers-Only responses.
...
- When receiving a Trailers-Only response, return the metadata as
trailing metadata instead of initial metadata.
- Send Trailers-Only response when we have no non-default initial metadata,
no message to send, and trailing metadata to send.
8 years ago
7897ae9308
auto-fix most of licenses
8 years ago
85a747ed06
better representative output
8 years ago
e76a0ecc03
Update hpack benchmarks for true binary
8 years ago
eb0e34f736
Convert everything to new encode API
8 years ago
83643bd5cb
Fix build on mac
8 years ago
61c317b624
Fix leak
8 years ago
0d04e0697b
Add a benchmark of base16 encoding
8 years ago
c945c12b47
Revert "Revert "Split bm_fullstack into pieces, reuse infrastructure across other microbenchmarks""
...
This reverts commit 0cb7fef406
8 years ago
0cb7fef406
Revert "Split bm_fullstack into pieces, reuse infrastructure across other microbenchmarks"
8 years ago
Esun Kim
Vijay Pai
Arjun Roy
Hope Casey-Allen
Soheil Hassas Yeganeh
Noah Eisen
ncteisen
Muxi Yan
Yash Tibrewal
Craig Tiller
yang-g
Mark D. Roth
Jan Tattermusch
Jan Tattermusch