[export]

Batch pending rpcs in client and server side before sending them down to tcp. If multiple rpcs get queued up into chaotic good, it should send them out over one tcp endpoint write.

----
DO NOT SUBMIT. This PR is for testing purposes only. [cl/627143927](http://cl/627143927)

PiperOrigin-RevId: 627143927

src/core/ext/transport/chaotic_good/chaotic_good_transport.h

@@ -67,6 +67,30 @@ class ChaoticGoodTransport : public RefCounted<ChaoticGoodTransport> {
         data_endpoint_.Write(std::move(buffers.data)));
   }
 
+  void SerializeFrameIntoBuffers(const FrameInterface& frame,
+                                 SliceBuffer& control_buffer,
+                                 SliceBuffer& data_buffer) {
+    auto buffers = frame.Serialize(&encoder_);
+    if (grpc_chaotic_good_trace.enabled()) {
+      gpr_log(GPR_INFO, "CHAOTIC_GOOD: WriteFrame to:%s %s",
+              ResolvedAddressToString(control_endpoint_.GetPeerAddress())
+                  .value_or("<<unknown peer address>>")
+                  .c_str(),
+              frame.ToString().c_str());
+    }
+    buffers.control.MoveFirstNBytesIntoSliceBuffer(buffers.control.Length(),
+                                                   control_buffer);
+    buffers.data.MoveFirstNBytesIntoSliceBuffer(buffers.data.Length(),
+                                                data_buffer);
+  }
+
+  auto WriteSerializedFrames(SliceBuffer& control_buffer,
+                             SliceBuffer& data_buffer) {
+    return TryJoin<absl::StatusOr>(
+        control_endpoint_.Write(std::move(control_buffer)),
+        data_endpoint_.Write(std::move(data_buffer)));
+  }
+
   // Read frame header and payloads for control and data portions of one frame.
   // Resolves to StatusOr<tuple<FrameHeader, BufferPair>>.
   auto ReadFrameBytes() {

src/core/ext/transport/chaotic_good/client_transport.cc

@@ -59,11 +59,17 @@ auto ChaoticGoodClientTransport::TransportWriteLoop(
     RefCountedPtr<ChaoticGoodTransport> transport) {
   return Loop([this, transport = std::move(transport)] {
     return TrySeq(
-        // Get next outgoing frame.
-        outgoing_frames_.Next(),
+        // Get all the next outgoing frames.
+        outgoing_frames_.AllNext(),
         // Serialize and write it out.
-        [transport = transport.get()](ClientFrame client_frame) {
-          return transport->WriteFrame(GetFrameInterface(client_frame));
+        [transport = transport.get()](std::vector<ClientFrame> client_frame) {
+          SliceBuffer control_buffer;
+          SliceBuffer data_buffer;
+          for (auto& frame : client_frame) {
+            transport->SerializeFrameIntoBuffers(GetFrameInterface(frame),
+                                                 control_buffer, data_buffer);
+          }
+          return transport->WriteSerializedFrames(control_buffer, data_buffer);
         },
         []() -> LoopCtl<absl::Status> {
           // The write failures will be caught in TrySeq and exit loop.

src/core/ext/transport/chaotic_good/server_transport.cc

@@ -54,11 +54,17 @@ auto ChaoticGoodServerTransport::TransportWriteLoop(
     RefCountedPtr<ChaoticGoodTransport> transport) {
   return Loop([this, transport = std::move(transport)] {
     return TrySeq(
-        // Get next outgoing frame.
-        outgoing_frames_.Next(),
+        // Get all the next outgoing frames.
+        outgoing_frames_.AllNext(),
         // Serialize and write it out.
-        [transport = transport.get()](ServerFrame client_frame) {
-          return transport->WriteFrame(GetFrameInterface(client_frame));
+        [transport = transport.get()](std::vector<ServerFrame> server_frame) {
+          SliceBuffer control_buffer;
+          SliceBuffer data_buffer;
+          for (auto& frame : server_frame) {
+            transport->SerializeFrameIntoBuffers(GetFrameInterface(frame),
+                                                 control_buffer, data_buffer);
+          }
+          return transport->WriteSerializedFrames(control_buffer, data_buffer);
         },
         []() -> LoopCtl<absl::Status> {
           // The write failures will be caught in TrySeq and exit loop.

src/core/lib/promise/mpsc.h

@@ -204,6 +204,27 @@ class MpscReceiver {
     };
   }
 
+  // Return a promise that will resolve to a vector containing all available
+  // items.
+  auto AllNext() {
+    return [this]() -> Poll<std::vector<T>> {
+      if (buffer_it_ == buffer_.end()) {
+        if (center_->PollReceiveBatch(buffer_)) {
+          buffer_it_ = buffer_.begin();
+        }
+      }
+      if (buffer_it_ != buffer_.end()) {
+        std::vector<T> dest;
+        buffer_it_ = buffer_.begin();
+        while (buffer_it_ != buffer_.end()) {
+          dest.push_back(std::move(*buffer_it_++));
+        }
+        return Poll<std::vector<T>>(std::move(dest));
+      }
+      return Pending{};
+    };
+  }
+
  private:
   // Received items. We move out of here one by one, but don't resize the
   // vector. Instead, when we run out of items, we poll the center for more -

test/core/promise/mpsc_test.cc

@@ -175,6 +175,30 @@ TEST(MpscTest, ImmediateSendWorks) {
   activity.Deactivate();
 }
 
+TEST(MpscTest, AllNextReceiveWorks) {
+  StrictMock<MockActivity> activity;
+  MpscReceiver<Payload> receiver(1);
+  MpscSender<Payload> sender = receiver.MakeSender();
+
+  EXPECT_EQ(sender.UnbufferedImmediateSend(MakePayload(1)), true);
+  EXPECT_EQ(sender.UnbufferedImmediateSend(MakePayload(2)), true);
+  EXPECT_EQ(sender.UnbufferedImmediateSend(MakePayload(3)), true);
+  EXPECT_EQ(sender.UnbufferedImmediateSend(MakePayload(4)), true);
+  EXPECT_EQ(sender.UnbufferedImmediateSend(MakePayload(5)), true);
+  EXPECT_EQ(sender.UnbufferedImmediateSend(MakePayload(6)), true);
+  EXPECT_EQ(sender.UnbufferedImmediateSend(MakePayload(7)), true);
+
+  activity.Activate();
+  auto payloads = NowOrNever(receiver.AllNext());
+  EXPECT_TRUE(payloads.has_value());
+  EXPECT_EQ(payloads.value().size(), 7);
+  int i = 1;
+  for (const auto& payload : *payloads) {
+    EXPECT_EQ(payload, MakePayload(i++));
+  }
+  activity.Deactivate();
+}
+
 }  // namespace
 }  // namespace grpc_core