Add support to multiple process groups by syncing across ranks (#151)

Summary:
Add support to multiple process groups by syncing across ranks.

Pull Request resolved: #151

Test Plan: /usr/local/fbcode/platform010/bin/mpirun -np 2 path-to/comm_replay.par --trace-path param_bench/fb/integration_tests/resnet-2gpu --trace-type et

Reviewed By: briancoutinho

Differential Revision: D60788539

Pulled By: shengfukevin

fbshipit-source-id: a33932be1bc81e4865d027b071fa855a347d857b

et_replay/comm/backend/base_backend.py

@@ -174,7 +174,6 @@ def alloc_ones(
             ipTensor = ipTensor * scaleFactor
         return ipTensor
 
-    @abstractmethod
     def noop(
         self,
         collectiveArgs: collectiveArgsHolder = None,

et_replay/comm/backend/pytorch_dist_backend.py

@@ -3,8 +3,10 @@
 # This source code is licensed under the MIT license found in the
 # LICENSE file in the root directory of this source tree.
 
+import json
 import logging
 import os
+
 from itertools import cycle
 from time import sleep
 from typing import List, Optional
@@ -1008,7 +1010,8 @@ def get_new_pg(self, group_ranks, backend):
                 ranks=group_ranks, backend=backend
             )
         else:
-            return dist.new_group(ranks=group_ranks, backend=backend)
+            pg = dist.new_group(ranks=group_ranks, backend=backend)
+            return pg
 
     def tensor_list_to_numpy(self, tensorList):
         if isinstance(tensorList, list):
@@ -1070,9 +1073,30 @@ def initialize_backend(
     def initialize_groups(self, backend="gloo"):
         groups = {}
         world_size = self.get_world_size()
+        global_rank = self.get_global_rank()
+
+        # sync pgs across ranks to fix hang with multiple comm groups
+        # because new_group() functions requires that all processes in the main group enter,
+        # even if they are not going to be members of the group.
+        # Assumption: pg_name is unique and consistent for all ranks
+        sync_store = dist.PrefixStore("pg_sync_r", self.tcp_store)
+        sync_store.set(str(global_rank), json.dumps(self.commsParams.groupRanks))
+        torch.distributed.barrier()
+        group_ranks_sync = self.commsParams.groupRanks.copy()
+        for i in range(self.get_world_size()):
+            if i == global_rank:
+                continue
+            json_data = sync_store.get(str(i))
+
+            # convert pg_id in json_data to int
+            pg_id2group_ranks = {
+                int(pg_id): rank for pg_id, rank in json.loads(json_data).items()
+            }
+
+            group_ranks_sync.update(pg_id2group_ranks)
 
         # create additional groups
-        for pg_id, group_ranks in self.commsParams.groupRanks.items():
+        for pg_id, group_ranks in dict(sorted(group_ranks_sync.items())).items():
             if (
                 len(group_ranks) > world_size
             ):  # this means that --auto-shrink is enabled, only use default pg
@@ -1084,11 +1108,9 @@ def initialize_groups(self, backend="gloo"):
                 pg = self.get_default_group()
             else:
                 pg = self.get_new_pg(group_ranks=group_ranks, backend=backend)
-                global_rank = self.get_global_rank()
-                if global_rank in group_ranks:
-                    logger.info(
-                        f"initialize_groups: Rank {global_rank} creates new group pg_id {pg_id} {pg} with {group_ranks}"
-                    )
+                logger.info(
+                    f"initialized_group: create new group, pg_id = {pg_id}, group_ranks = {group_ranks}"
+                )
             groups[pg_id] = pg
 
         # if additional groups are created, overwrite the default groups list

et_replay/tools/comm_replay.py

@@ -31,7 +31,8 @@
 from et_replay.comm.param_profile import paramProfile, paramTimer
 
 try:
-    from trainer_iteration_wrapper import setTrainingIteration  # @manual
+    # pyre-ignore[21]:
+    from trainer_iteration_wrapper import setTrainingIteration
 except ImportError:
     pass
 
@@ -89,6 +90,8 @@ def writeCommDetails(commsTracePerf: List, rank: int, folder: str = "./") -> Non
             json.dump(commsTracePerf, write_file, indent=2)
 
 
+# pyre-ignore[13]: lint complained about self.backendFuncs is never initlized.
+#                  it is initialized in initBackend
 class commsTraceReplayBench(paramCommsBench):
     """
     A class to replay and benchmark generated traces for collective communications.