Benjamin Mahler created MESOS-755:
-------------------------------------
Summary: Using the soft limit with memory thresholds has different
OOM semantics than simply setting the hard limit.
Key: MESOS-755
URL: https://issues.apache.org/jira/browse/MESOS-755
Project: Mesos
Issue Type: Bug
Affects Versions: 0.15.0, 0.14.1
Reporter: Benjamin Mahler
In order to fix MESOS-662, we enabled the OOM killer.
In addition, instead of setting the memory hard limit (memory.limit_in_bytes),
we set the soft limit (memory.soft_limit_in_bytes) to the requested amount of
memory and set the hard limit higher by a fixed amount. Once the soft limit is
reached, this triggers a memory threshold notification at which point we
capture the memory.stat information and treat the executor as having OOMed.
We've seen reports from users that this is not behaving the same as simply
setting the hard limit. In particular, we've seen the file cache not being
purged by the kernel upon hitting the soft limit (the kernel documentation only
states that action is taken with the soft limit in the presence of system wide
memory pressure). However, it was not clear over email and in the review as to
the extent of which the hard limit and soft limit are treated differently in
terms of purging cached memory: https://reviews.apache.org/r/14043/
--
This message was sent by Atlassian JIRA
(v6.1#6144)
