Increase TCP listen queue size from 64 to 1024 #10268
Closed
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
adamlerer
requested review from
apaszke,
colesbury,
ezyang,
gchanan,
soumith and
zdevito
as code owners
adamlerer
force-pushed
the
export-D9182216
branch
from
96eb5f1 to
f4d3240
Compare
Summary: Pull Request resolved: pytorch#10268 Running torch.distributed.init_process_group fails with more than ~64 processes, with various errors like connection refused or connection reset by peer. After some digging, it looks like the root cause is that all workers have to connect to master via TCP (both in Zeus init and in DataChannelTCP - look for `connect()`), and the listening socket only has a backlog of 64. I increased the backlog to 1024, that seems like enough for reasonable purposes (the hard limit is 65535 in /proc/sys/net/core/somaxconn). There's probably a more correct way to do this that involves retries when connection is refused. Differential Revision: D9182216 fbshipit-source-id: d57c37c20fc1fdbcc24b5064a29174ca580c43d6
adamlerer
force-pushed
the
export-D9182216
branch
from
f4d3240 to
6dbd552
Compare
soumith
approved these changes
Aug 6, 2018
PenghuiCheng
pushed a commit
to PenghuiCheng/pytorch
that referenced
this pull request
Aug 10, 2018
Summary: Pull Request resolved: pytorch#10268 Running torch.distributed.init_process_group fails with more than ~64 processes, with various errors like connection refused or connection reset by peer. After some digging, it looks like the root cause is that all workers have to connect to master via TCP (both in Zeus init and in DataChannelTCP - look for `connect()`), and the listening socket only has a backlog of 64. I increased the backlog to 1024, that seems like enough for reasonable purposes (the hard limit is 65535 in /proc/sys/net/core/somaxconn). There's probably a more correct way to do this that involves retries when connection is refused. Reviewed By: soumith Differential Revision: D9182216 fbshipit-source-id: 2f71c4995841db26c670cec344f1e3c7a80a7936
goodlux
pushed a commit
to goodlux/pytorch
that referenced
this pull request
Aug 15, 2018
Summary: Pull Request resolved: pytorch#10268 Running torch.distributed.init_process_group fails with more than ~64 processes, with various errors like connection refused or connection reset by peer. After some digging, it looks like the root cause is that all workers have to connect to master via TCP (both in Zeus init and in DataChannelTCP - look for `connect()`), and the listening socket only has a backlog of 64. I increased the backlog to 1024, that seems like enough for reasonable purposes (the hard limit is 65535 in /proc/sys/net/core/somaxconn). There's probably a more correct way to do this that involves retries when connection is refused. Reviewed By: soumith Differential Revision: D9182216 fbshipit-source-id: 2f71c4995841db26c670cec344f1e3c7a80a7936
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary:
Running torch.distributed.init_process_group fails with more than ~64 processes, with various errors like connection refused or connection reset by peer. After some digging, it looks like the root cause is that all workers have to connect to master via TCP (both in Zeus init and in DataChannelTCP - look for
connect()), and the listening socket only has a backlog of 64.I increased the backlog to 1024, that seems like enough for reasonable purposes (the hard limit is 65535 in /proc/sys/net/core/somaxconn). There's probably a more correct way to do this that involves retries when connection is refused.
Differential Revision: D9182216