Merge pull request #9272 from ctiller/perfnotes

Start some performance notes

1 files changed, 29 insertions, 0 deletions

diff --git a/doc/cpp/perf_notes.md b/doc/cpp/perf_notes.md
new file mode 100644
index 0000000000..805ead3638
--- /dev/null
+++ b/doc/cpp/perf_notes.md
@@ -0,0 +1,29 @@
+# C++ Performance Notes
+
+## Streaming write buffering
+
+Generally, each write operation (Write(), WritesDone()) implies a syscall.
+gRPC will try to batch together separate write operations from different
+threads, but currently cannot automatically infer batching in a single stream.
+
+If message k+1 in a stream does not rely on responses from message k, it's
+possible to enable write batching by passing a WriteOptions argument to Write
+with the buffer_hint set:
+
+```c++
+stream_writer->Write(message, WriteOptions().set_buffer_hint());
+```
+
+The write will be buffered until one of the following is true:
+- the per-stream buffer is filled (controllable with the channel argument
+  GRPC_ARG_HTTP2_WRITE_BUFFER_SIZE) - this prevents infinite buffering leading
+  to OOM
+- a subsequent Write without buffer_hint set is posted
+- the call is finished for writing (WritesDone() called on the client,
+  or Finish() called on an async server stream, or the service handler returns
+  for a sync server stream)
+
+## Completion Queues and Threading in the Async API
+
+Right now, the best performance trade-off is having numcpu's threads and one
+completion queue per thread.