Centralize listener connection support #613
Merged
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
…were failing to start due to connection failures of the opal/pmix support. Root cause was that (a) we were setting the client socket to non-blocking before calling connect, and (b) the server was using the event library to harvest the accepts, and also did the handshake while in that event. So the server would backup beyond the connection backlog limit, and we would fail. Changing the client to leave its socket as blocking during the connect doesn't solve the problem by itself - you also have to introduce a sleep delay once the backlog is hit to avoid simply machine-gunning your way thru retries. This gets somewhat difficult to adjust as you don't want to unnecessarily prolong startup time. We've solved this before by adding a listening thread that simply reaps accepts and shoves them into the event library for subsequent processing. This would resolve the problem, but meant yet another daemon-level thread. So I centralized the listening thread support and let multiple elements register listeners on it. Thus, each daemon now has a single listening thread that reaps accepts from multiple sources - for now, the orte/pmix server and the oob/usock support are using it. I'll add in the oob/tcp component later. This still didn't fully resolve the SMP problem, especially on coprocessor cards (e.g., KNC). Removing the shared memory dstore support helped further improve the behavior - it looks like there is some kind of memory paging issue there that needs further understanding. Given that the shared memory support was about to be lost when I bring over the PMIx integration (until it is restored in that library), it seemed like a reasonable thing to just remove it at this point.
|
Sent from
|
Refer to this link for build results (access rights to CI server needed): |
rhc54
pushed a commit
that referenced
this pull request
May 29, 2015
Centralize listener connection support
jsquyres
pushed a commit
to jsquyres/ompi
that referenced
this pull request
Sep 19, 2016
…I_Ibarrier man: fix a typo in MPI_Ibarrier C prototype
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 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.
Fat SMPs (i.e., systems with nodes containing large numbers of cpus) were failing to start due to connection failures of the opal/pmix support. Root cause was that (a) we were setting the client socket to non-blocking before calling connect, and (b) the server was using the event library to harvest the accepts, and also did the handshake while in that event. So the server would backup beyond the connection backlog limit, and we would fail.
Changing the client to leave its socket as blocking during the connect doesn't solve the problem by itself - you also have to introduce a sleep delay once the backlog is hit to avoid simply machine-gunning your way thru retries. This gets somewhat difficult to adjust as you don't want to unnecessarily prolong startup time.
We've solved this before by adding a listening thread that simply reaps accepts and shoves them into the event library for subsequent processing. This would resolve the problem, but meant yet another daemon-level thread. So I centralized the listening thread support and let multiple elements register listeners on it. Thus, each daemon now has a single listening thread that reaps accepts from multiple sources - for now, the orte/pmix server and the oob/usock support are using it. I'll add in the oob/tcp component later.
This still didn't fully resolve the SMP problem, especially on coprocessor cards (e.g., KNC). Removing the shared memory dstore support helped further improve the behavior - it looks like there is some kind of memory paging issue there that needs further understanding. Given that the shared memory support was about to be lost when I bring over the PMIx integration (until it is restored in that library), it seemed like a reasonable thing to just remove it at this point.