@ -10,9 +10,9 @@ Polling engine component was created for the following reasons:
- gRPC code deals with a bunch of file descriptors on which events like descriptor being readable/writable/error have to be monitored
- gRPC code knows the actions to perform when such events happen
- For example:
- `grpc_endpoint` code calls recvmsg call when the fd is readable and sendmsg call when the fd is writable
- ` tcp_client` connect code issues async connect and finishes creating the client once the fd is writable (i.e when the connect actually finished)
- gRPC needed some component that can "efficiently" to the above operations __using the threads provided by the applications (i.e not create any new threads)__. Also by "efficiently" we mean optimized for latency and throughput
- `grpc_endpoint` code calls `recvmsg` call when the fd is readable and `sendmsg` call when the fd is writable
- ` tcp_client` connect code issues async `connect` and finishes creating the client once the fd is writable (i.e when the `connect` actually finished)
- gRPC needed some component that can "efficiently" do the above operations __using the threads provided by the applications (i.e., not create any new threads)__. Also by "efficiently" we mean optimized for latency and throughput
## Polling Engine Implementations in gRPC
@ -20,15 +20,15 @@ There are multiple polling engine implementations depending on the OS and the OS
- Linux:
- **epollex** (default but requires kernel version >= 4.5),
- epoll1 (If epollex is not available and glibc version >= 2.9)
- epollsig (if epollex, epoll1 are unavailable AND Kernel has epoll support)
- poll (if kernel NOT have epoll support)
- Mac: **poll** (default), poll-cv
- **`epollex`** (default but requires kernel version >= 4.5),
- `epoll1` (If `epollex` is not available and glibc version >= 2.9)
@ -36,8 +36,8 @@ There are multiple polling engine implementations depending on the OS and the OS
The following are the **Opaque** structures exposed by Polling Engine interface (NOTE: Different polling engine implementations have different definitions of these structures)
- **grpc_fd:** Structure representing a file descriptor
- **grpc_pollset:** A set of one or more grpc_fds that are ‘polled’ for readable/writable/error events. One grpc_fd can be in multiple grpc_pollsets
- **grpc_pollset_worker:** Structure representing a ‘polling thread’ - more specifically, the thread that calls grpc_pollset_work() API
- **grpc_pollset:** A set of one or more grpc_fds that are ‘polled’ for readable/writable/error events. One grpc_fd can be in multiple `grpc_pollset`s
- **grpc_pollset_worker:** Structure representing a ‘polling thread’ - more specifically, the thread that calls `grpc_pollset_work()` API
- **grpc_pollset_set:** A group of `grpc_fds`, `grpc_pollsets` and `grpc_pollset_sets` (yes, a `grpc_pollset_set` can contain other `grpc_pollset_sets`)
### Polling engine API
@ -45,7 +45,7 @@ The following are the **Opaque** structures exposed by Polling Engine interface
- Register a closure to be called when the fd becomes readable/writable or has an error (In grpc parlance, we refer to this act as “arming the fd”)
- Register a [closure](https://github.com/grpc/grpc/blob/v1.15.1/src/core/lib/iomgr/closure.h#L67) to be called when the fd becomes readable/writable or has an error (In grpc parlance, we refer to this act as “arming the fd”)
- The closure is called exactly once per event. I.e once the fd becomes readable (or writable or error), the closure is fired and the fd is ‘unarmed’. To be notified again, the fd has to be armed again.
- **grpc_fd_shutdown**
@ -56,7 +56,7 @@ The following are the **Opaque** structures exposed by Polling Engine interface
- Release the `grpc_fd` structure and call `on_done` closure when the operation is complete
- If `release_fd` is set to `nullptr`, then `close()` the underlying fd as well. If not, put the underlying fd in `release_fd` (and do not call `close()`)
- release_fd set to non-null in cases where the underlying fd is NOT owned by grpc core (like for example the fds used by C-Ares DNS resolver )
- `release_fd` set to non-null in cases where the underlying fd is NOT owned by grpc core (like for example the fds used by C-Ares DNS resolver )
#### grpc_pollset
@ -68,10 +68,10 @@ The following are the **Opaque** structures exposed by Polling Engine interface
> **NOTE**: `grpc_pollset_work()` requires the pollset mutex to be locked before calling it. Shortly after calling `grpc_pollset_work()`, the function populates the `*worker` pointer (among other things) and releases the mutex. Once `grpc_pollset_work()` returns, the `*worker` pointer is **invalid** and should not be used anymore. See the code in `completion_queue.cc` to see how this is used.
- Poll the fds in the pollset for events AND return when ONE of the following is true:
- Poll the fds in the pollset for events AND return when ANY of the following is true:
- Deadline expired
- Some fds in the pollset were found to be readable/writable/error and those associated closures were ‘scheduled’ (but not necessarily executed)
- worker is “kicked” (see grpc_pollset_kick for more details)
- worker is “kicked” (see `grpc_pollset_kick` for more details)
- The logic to choose a designated poller is quite complicated. Pollsets are internally sharded into what are called `pollset_neighborhood` (a structure internal to `epoll1` polling engine implementation). `grpc_pollset_workers` that call `grpc_pollset_work` on a given pollset are all queued in a linked-list against the `grpc_pollset`.
- The logic to choose a designated poller is quite complicated. Pollsets are internally sharded into what are called `pollset_neighborhood` (a structure internal to `epoll1` polling engine implementation). `grpc_pollset_workers` that call `grpc_pollset_work` on a given pollset are all queued in a linked-list against the `grpc_pollset`. The head of the linked list is called "root worker"
- There are as many neighborhoods as the number of cores. A pollset is put in a neighborhood based on the CPU core, the root worker thread (i.e the head of the linked-list of workers queued against the pollset). The whole idea here is that when choosing the next designated poller, we make a best-effort attempt to pick a worker that is NOT running on the same core. This way, we reduce the probability of the current thread being pre-empted by the CPU scheduler.
- See `begin_worker()` function in `src/core/lib/iomgr/ev_epoll1_posix.cc` to see how a designated poller is chosen. Similarly `end_worker()` function is called by the worker that was just out of epoll_wait() and will have to choose a new designated poller)
- There are as many neighborhoods as the number of cores. A pollset is put in a neighborhood based on the CPU core of the root worker thread. When picking the next designated poller, we always try to find another worker on the current pollset. If there are no more workers in the current pollset, a `pollset_neighborhood` listed is scanned to pick the next pollset and worker that could be the new designated poller.
- NOTE: There is room to tune this implementation. All we really need is good way to maintain a list of `grpc_pollset_workers` with a way to group them per-pollset (needed to implement `grpc_pollset_kick` semantics) and a way randomly select a new designated poller
- See [`begin_worker()`](https://github.com/grpc/grpc/blob/v1.15.1/src/core/lib/iomgr/ev_epoll1_linux.cc#L729) function to see how a designated poller is chosen. Similarly [`end_worker()`](https://github.com/grpc/grpc/blob/v1.15.1/src/core/lib/iomgr/ev_epoll1_linux.cc#L916) function is called by the worker that was just out of `epoll_wait()` and will have to choose a new designated poller)
### epollex
@ -130,17 +130,17 @@ Code at `src/core/lib/iomgr/ev_epollex_posix.cc`
- FDs are added to multiple epollsets with EPOLLEXCLUSIVE flag. This prevents multiple worker threads from waking up from polling whenever the fd is readable/writable
- A few conclusions:
- A few observations:
- If multiple pollsets are pointing to the same Pollable, then the pollable MUST be either empty or of type PO_FD (i.e single-fd)
- A multi-pollable has one-and-only-one incoming link from a Pollset
- The same FD can be in multiple pollables (even if one of the pollables is of type PO_FD)
- There cannot be two Pollables of type PO_FD for the same fd
- If multiple pollsets are pointing to the same `Pollable`, then the `pollable` MUST be either empty or of type `PO_FD` (i.e single-fd)
- A multi-pollable has one-and-only-one incoming link from a pollset
- The same FD can be in multiple `Pollable`s (even if one of the `Pollable`s is of type PO_FD)
- There cannot be two `Pollable`s of type PO_FD for the same fd
- Why do we need Pollalbe of type PO_FD and Empty pollable ?
- Why do we need `Pollable` of type PO_FD and PO_EMTPY ?
- The main reason is the Sync client API
- We create one completion queue per call (therefore one pollset per call). If we didn’t have PO_EMPTY and PO_FD type pollables, then every call on a given channel will effectively have to create a pollable (and hence an epollset). Thats a lot of epoll fd create/delete calls
- With these new types of pollables, all pollsets (corresponding to the new per-call completion queue) will initially point to PO_EMPTY global epollset. Then once the sub-channel fd is added to the pollset, the pollset will point to the Pollable of type PO_FD containing just that fd (i.e
- We create one new completion queue per call. If we didn’t have PO_EMPTY and PO_FD type pollables, then every call on a given channel will effectively have to create a `Pollable` and hence an epollset. This is because every completion queue automatically creates a pollset and the channel fd will have to be put in that pollset. This clearly requires an epollset to put that fd. Creating an epollset per call (even if we delete the epollset once the call is completed) would mean a lot of sys calls to create/delete epoll fds. This is clearly not a good idea.
- With these new types of `Pollable`s, all pollsets (corresponding to the new per-call completion queue) will initially point to PO_EMPTY global epollset. Then once the channel fd is added to the pollset, the pollset will point to the `Pollable` of type PO_FD containing just that fd (i.e it will reuse the existing `Pollable`). This way, the epoll fd creation/deletion churn is avoided.
### Other polling engine implementations (poll and windows polling engine)
Sree Kuchibhotla
6 years ago