[
https://issues.apache.org/jira/browse/MESOS-8038?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16556308#comment-16556308
]
Zhitao Li commented on MESOS-8038:
----------------------------------
Some update:
We have another episode on this issue. Our setup is our custom framework +
command executor on GPU in a cluster with GPU only machines and GPU only tasks,
running Mesos 1.5.0.
One error log pattern that 100% correlates in our environment:
bq. E0724 01:14:15.203124 10883 slave.cpp:5798] Termination of executor
'3e213d20-ed99-4196-bd26-12560423c5fd-151-1' of framework
8e2c0e03-3147-442c-9abe-b370aad201cd-0000 failed: Failed to kill all processes
in the container: Timed out after 1mins
Based on [this
TODO|https://github.com/apache/mesos/blob/1.5.0/src/slave/containerizer/mesos/containerizer.cpp#L2504],
containerizer would not call into nvidia/gpu isolator's cleanup method, thus
the gpu resource would "leaked" until next agent restart, which matches our
observation.
Based on the error logs, we are looking at the following:
1. Why would the termination takes longer than the
[DESTROY_TIMEOUT|https://github.com/apache/mesos/blob/a86ff8c36532f97b6eb6b44c6f871de24afbcc4d/src/linux/cgroups.hpp#L44]?
2. Should this be configurable?
3. should the early return in above case really happen?
[~jasonlai] and I might discuss this issue in next containerizer WG.
> Launching GPU task sporadically fails.
> --------------------------------------
>
> Key: MESOS-8038
> URL: https://issues.apache.org/jira/browse/MESOS-8038
> Project: Mesos
> Issue Type: Bug
> Components: allocation, containerization, gpu
> Affects Versions: 1.4.0
> Reporter: Sai Teja Ranuva
> Assignee: Zhitao Li
> Priority: Critical
> Attachments: mesos-master.log, mesos-slave.INFO.log
>
>
> I was running a job which uses GPUs. It runs fine most of the time.
> But occasionally I see the following message in the mesos log.
> "Collect failed: Requested 1 but only 0 available"
> Followed by executor getting killed and the tasks getting lost. This happens
> even before the the job starts. A little search in the code base points me to
> something related to GPU resource being the probable cause.
> There is no deterministic way that this can be reproduced. It happens
> occasionally.
> I have attached the slave log for the issue.
> Using 1.4.0 Mesos Master and 1.4.0 Mesos Slave.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)