Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

fix port conflict in parallel dygraph mode #28497

Merged
merged 2 commits into from
Nov 16, 2020
Merged

fix port conflict in parallel dygraph mode #28497

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
f485474
fix nccl init failed in dygraph mode;test=develop
danleifeng Nov 9, 2020
bfdaffb
change order for wait_server_ready;test=develop
danleifeng Nov 11, 2020
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
38 changes: 21 additions & 17 deletions python/paddle/distributed/parallel.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -125,7 +125,7 @@ def _check_var_exists(var_name):
if ParallelEnv().world_size < 2:
return

# 3: init gloo context
# 3: init gloo context (step 1: httpsever start)
ep_rank_0 = ParallelEnv().trainer_endpoints[0].split(":")
ep_rank = ParallelEnv().trainer_endpoints[ParallelEnv().rank].split(":")
manager = Manager()
Expand All @@ -140,21 +140,6 @@ def _check_var_exists(var_name):
http_server.start()
wait_server_ready([ParallelEnv().trainer_endpoints[0]])
Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This line can be moved to line #169 (just after the nccl initialization) to decrease the wait time.


gloo_strategy = core.GlooParallelStrategy()
gloo_strategy.rank = ParallelEnv().rank
gloo_strategy.rank_num = ParallelEnv().world_size
gloo_strategy.ip_address = ep_rank_0[0]
gloo_strategy.ip_port = int(ep_rank_0[1])
default_init_timeout_seconds = 3600
default_run_timeout_seconds = 9999999
gloo_strategy.init_seconds = default_init_timeout_seconds
gloo_strategy.run_seconds = default_run_timeout_seconds
gloo = core.GlooParallelContext(gloo_strategy)
gloo.init()
if ParallelEnv().rank == 0:
http_server_d["running"] = False
http_server.join()

# 4. init NCCL ParallelStrategy
strategy = ParallelStrategy()
if parallel_helper._is_parallel_ctx_initialized():
Expand All @@ -165,7 +150,7 @@ def _check_var_exists(var_name):
strategy.current_endpoint = ParallelEnv().current_endpoint

# NOTE(chenweihang): [ why config global place here? ]
# the dygraph mode will be set to default mode,
# the dygraph mode will be set to default mode,
# users will not call `dygraph.guard` or `enable_dygraph`
# directly, if they want to switch default place,
# they need to call a function to change default place,
Expand All @@ -177,6 +162,25 @@ def _check_var_exists(var_name):
parallel_helper._set_parallel_ctx(core.NCCLParallelContext(strategy, place))
parallel_helper._init_parallel_ctx()

# 5: init gloo context (step 2: gloo init)
# dividing init_gloo into two part beacause nccl and gloo
# are separately looking for free ports which sometimes
# leads to port-conflict.
gloo_strategy = core.GlooParallelStrategy()
gloo_strategy.rank = ParallelEnv().rank
gloo_strategy.rank_num = ParallelEnv().world_size
gloo_strategy.ip_address = ep_rank_0[0]
gloo_strategy.ip_port = int(ep_rank_0[1])
default_init_timeout_seconds = 3600
default_run_timeout_seconds = 9999999
gloo_strategy.init_seconds = default_init_timeout_seconds
gloo_strategy.run_seconds = default_run_timeout_seconds
gloo = core.GlooParallelContext(gloo_strategy)
gloo.init()
if ParallelEnv().rank == 0:
http_server_d["running"] = False
http_server.join()


def get_rank():
"""
Expand Down