Reuse channel to send data. (#141)

Before this PR, we'll create a new channel object when sending data. We reuse that in this PR to make code clean.

Also, we did a benchmark from @fengsp(Thanks for the contribution) for this PR, and the result is not too good.

The baseline(before this PR) is:
```
num calls: 10000
total time (ms) =  79241.70279502869
per task overhead (ms) = 7.924171590805054
```

and the current result after this PR is:
```
num calls: 10000
total time (ms) =  78388.72385025024
per task overhead (ms) = 7.83887369632721
```

---------

Signed-off-by: Qing Wang <kingchin1218@gmail.com>

benchmarks/many_tiny_tasks_benchmark.py

@@ -0,0 +1,65 @@
+# Copyright 2023 The RayFed Team
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import ray
+import time
+import sys
+import fed
+
+
+@fed.remote
+class MyActor:
+    def run(self):
+        return None
+
+
+@fed.remote
+class Aggregator:
+    def aggr(self, val1, val2):
+        return None
+
+
+def main(party):
+    ray.init(address='local')
+
+    cluster = {
+        'alice': {'address': '127.0.0.1:11010'},
+        'bob': {'address': '127.0.0.1:11011'},
+    }
+    fed.init(cluster=cluster, party=party)
+
+    actor_alice = MyActor.party("alice").remote()
+    actor_bob = MyActor.party("bob").remote()
+    aggregator = Aggregator.party("alice").remote()
+
+    start = time.time()
+    num_calls = 10000
+    for i in range(num_calls):
+        val_alice = actor_alice.run.remote()
+        val_bob = actor_bob.run.remote()
+        sum_val_obj = aggregator.aggr.remote(val_alice, val_bob)
+        fed.get(sum_val_obj)
+        if i % 100 == 0:
+            print(f"Running {i}th call")
+    print(f"num calls: {num_calls}")
+    print("total time (ms) = ", (time.time() - start)*1000)
+    print("per task overhead (ms) =", (time.time() - start)*1000/num_calls)
+
+    fed.shutdown()
+    ray.shutdown()
+
+
+if __name__ == "__main__":
+    assert len(sys.argv) == 2, 'Please run this script with party.'
+    main(sys.argv[1])

fed/proxy/barriers.py

@@ -124,65 +124,30 @@ async def _run_grpc_server(
 
 
 async def send_data_grpc(
-    dest,
     data,
+    stub,
     upstream_seq_id,
     downstream_seq_id,
     metadata=None,
-    tls_config=None,
-    retry_policy=None,
-    grpc_options=None
 ):
-    grpc_options = get_grpc_options(retry_policy=retry_policy) if \
-                    grpc_options is None else fed_utils.dict2tuple(grpc_options)
-    tls_enabled = fed_utils.tls_enabled(tls_config)
     cluster_config = fed_config.get_cluster_config()
-    metadata = fed_utils.dict2tuple(metadata)
-    if tls_enabled:
-        ca_cert, private_key, cert_chain = fed_utils.load_cert_config(tls_config)
-        credentials = grpc.ssl_channel_credentials(
-            certificate_chain=cert_chain,
-            private_key=private_key,
-            root_certificates=ca_cert,
-        )
-
-        async with grpc.aio.secure_channel(
-            dest,
-            credentials,
-            options=grpc_options,
-        ) as channel:
-            stub = fed_pb2_grpc.GrpcServiceStub(channel)
-            data = cloudpickle.dumps(data)
-            request = fed_pb2.SendDataRequest(
-                data=data,
-                upstream_seq_id=str(upstream_seq_id),
-                downstream_seq_id=str(downstream_seq_id),
-            )
-            # wait for downstream's reply
-            response = await stub.SendData(
-                request, metadata=metadata, timeout=cluster_config.cross_silo_timeout)
-            logger.debug(
-                f'Received data response from seq_id {downstream_seq_id}, '
-                f'result: {response.result}.'
-            )
-            return response.result
-    else:
-        async with grpc.aio.insecure_channel(dest, options=grpc_options) as channel:
-            stub = fed_pb2_grpc.GrpcServiceStub(channel)
-            data = cloudpickle.dumps(data)
-            request = fed_pb2.SendDataRequest(
-                data=data,
-                upstream_seq_id=str(upstream_seq_id),
-                downstream_seq_id=str(downstream_seq_id),
-            )
-            # wait for downstream's reply
-            response = await stub.SendData(
-                request, metadata=metadata, timeout=cluster_config.cross_silo_timeout)
-            logger.debug(
-                f'Received data response from seq_id {downstream_seq_id} '
-                f'result: {response.result}.'
-            )
-            return response.result
+    data = cloudpickle.dumps(data)
+    request = fed_pb2.SendDataRequest(
+        data=data,
+        upstream_seq_id=str(upstream_seq_id),
+        downstream_seq_id=str(downstream_seq_id),
+    )
+    # Waiting for the reply from downstream.
+    response = await stub.SendData(
+        request,
+        metadata=fed_utils.dict2tuple(metadata),
+        timeout=cluster_config.cross_silo_timeout,
+    )
+    logger.debug(
+        f'Received data response from seq_id {downstream_seq_id}, '
+        f'result: {response.result}.'
+    )
+    return response.result
 
 
 @ray.remote
@@ -207,6 +172,8 @@ def __init__(
         self._tls_config = tls_config
         self.retry_policy = retry_policy
         self._grpc_metadata = fed_config.get_job_config().grpc_metadata
+        # Mapping the destination party name to the reused client stub.
+        self._stubs = {}
         cluster_config = fed_config.get_cluster_config()
         set_max_message_length(cluster_config.cross_silo_messages_max_size)
 
@@ -235,15 +202,33 @@ async def send(
         dest_addr = self._cluster[dest_party]['address']
         dest_party_grpc_config = self.setup_grpc_config(dest_party)
         try:
+            tls_enabled = fed_utils.tls_enabled(self._tls_config)
+            grpc_options = dest_party_grpc_config['grpc_options']
+            grpc_options = get_grpc_options(retry_policy=self.retry_policy) if \
+                grpc_options is None else fed_utils.dict2tuple(grpc_options)
+
+            if dest_party not in self._stubs:
+                if tls_enabled:
+                    ca_cert, private_key, cert_chain = fed_utils.load_cert_config(
+                        self._tls_config)
+                    credentials = grpc.ssl_channel_credentials(
+                        certificate_chain=cert_chain,
+                        private_key=private_key,
+                        root_certificates=ca_cert,
+                    )
+                    channel = grpc.aio.secure_channel(
+                        dest_addr, credentials, options=grpc_options)
+                else:
+                    channel = grpc.aio.insecure_channel(dest_addr, options=grpc_options)
+                stub = fed_pb2_grpc.GrpcServiceStub(channel)
+                self._stubs[dest_party] = stub
+
             response = await send_data_grpc(
-                dest=dest_addr,
                 data=data,
+                stub=self._stubs[dest_party],
                 upstream_seq_id=upstream_seq_id,
                 downstream_seq_id=downstream_seq_id,
                 metadata=dest_party_grpc_config['grpc_metadata'],
-                tls_config=self._tls_config,
-                retry_policy=self.retry_policy,
-                grpc_options=dest_party_grpc_config['grpc_options']
             )
         except Exception as e:
             logger.error(f'Failed to {send_log_msg}, error: {e}')