-
Notifications
You must be signed in to change notification settings - Fork 939
v4.0.x: Fix a couple managed allocation issues. #8352
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.
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
Conversation
If the user asks for a hostfile/hostlist inside of a managed allocation, make sure that rmaps filters these and maps processes based on them. Otherwise, it can result in inconsistent mappings across root and compute nodes if the user orders their hostfile differently than the resource manager. Signed-off-by: Austen Lauria <[email protected]>
For example: $. bsub -n 40 -m "node1 node2" mpirun -np 6 -host node1:2,node2:4 hostname would not map two hostname processes to node1 and four to node2. Instead, it would still think that each node1 and node2 had (for example) 20 cpu resources, and map accordingly. Signed-off-by: Austen Lauria <[email protected]>
|
Is this fixing a customer observed issue or was it found with dev testing? Does this code affect other resource managers other than LSF? If so, which ones, and could we ask them to test this before merging to release branch? Also, is it possible that this could break existing jobs submission scripts if they were relying on the older (presumably broken) functionality? Finally, could you also please make a v4.1 version of this as it also relies on orte launcher, and presumably would have a similar bug? |
|
Commit 1 was a customer reported issue who uses LSF. I only tested these fixes under LSF - I think we would need others from the community to test other RM's. Commit 1 fixes a bug where the ranks are mapped inconsistently across orted's, resulting in PMIx failures and other weirdness (such as multiple proc's being assigned the same rank, and hangs in MPI_Finalize()). Commit 2 fixes a bug as described in the message. This was not customer reported, but I found it in my testing of the fix. I plan to make a 4.1 PR once this is merged. I figure if people have changes/suggestions, I'd rather just fix it once than twice. |
|
From reviewing the code this could impact Slurm as well - I don't see others, but am not 100% on that.. @gpaulsen do you know who to tag for slurm/other testing? |
|
We can ask on Tuesday. @rhc54 might know if it's worth hand testing? |
rhc54
left a comment
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This will, of course, be seen by other RMs and not just LSF. However, I don't see any harm in fixing the bug. It only impacts if they specify -host or -hostfile, which isn't that common a case.
|
@awlauria Please upstream to PRRTE as well - if you don't have time, let me know so we can ensure it gets there. |
|
Thanks @rhc54 . Ported over: openpmix/prrte#718 |
|
v4.1.x: #8355 |
If the user asks for a hostfile/hostlist inside of a managed allocation,
make sure that rmaps filters these and maps processes based on them. Otherwise,
it can result in inconsistent mappings across root and compute nodes if the
user orders their hostfile differently than the resource manager.
For example:
$. bsub -n 40 -m "node1 node2" mpirun -np 6 -host node1:2,node2:4 hostname
would not map two hostname processes to node1 and four to node2.
Instead, it would still think that each node1
and node2 had (for example) 20 cpu resources, and map accordingly.