[jira] [Comment Edited] (MESOS-8418) mesos-agent high cpu usage because of numerous /proc/mounts reads

["Stephan Erb (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/MESOS-8418?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16543743#comment-16543743 ] 

Stephan Erb edited comment on MESOS-8418 at 7/13/18 9:32 PM:
-------------------------------------------------------------

I have attached a profile [^mesos-agent.stacks.gz] gathered on a host with:
 * ~140 tasks, all running docker images via the Mesos containerizer
 * relevant isolators {{cgroups/cpu,cgroups/mem,filesystem/linux,docker/runtime,...'}}
 * agent monitoring endpoints {{/slave(1)/monitor/statistic}} and {{/slave(1)/state}} are scraped every 15s. In total, a scrape takes between 5-13sec
 * potentially related agent settings:
 ** --oversubscribed_resources_interval=45secs
 ** --qos_correction_interval_min=45secs
 ** --cgroups_cpu_enable_pids_and_tids_count

The profile confirms that reading a mount table takes significant time, and is also done multiple times in short succession:

!mesos-agent-flamegraph.png|width=800,height=450!


was (Author: stephanerb):


I have attached a profile [^mesos-agent.stacks.gz] gathered on a host with:
 * ~140 tasks, all running docker images via the Mesos containerizer
 * relevant isolators {{cgroups/cpu,cgroups/mem,filesystem/linux,docker/runtime,...'}}
 * agent monitoring endpoints {{/slave(1)/monitor/statistic}} and {{/slave(1)/state}} are scraped every 15s. In total, a scrape takes between 5-13sec
 * potentially related agent settings:
 ** --oversubscribed_resources_interval=45secs
 ** --qos_correction_interval_min=45secs
 ** --cgroups_cpu_enable_pids_and_tids_count

The profile confirms that mount table reads take a significant amount of time

!mesos-agent-flamegraph.png|width=800,height=450!

> mesos-agent high cpu usage because of numerous /proc/mounts reads
> -----------------------------------------------------------------
>
>                 Key: MESOS-8418
>                 URL: https://issues.apache.org/jira/browse/MESOS-8418
>             Project: Mesos
>          Issue Type: Improvement
>          Components: agent, containerization
>            Reporter: Stéphane Cottin
>            Priority: Major
>              Labels: containerizer, performance
>         Attachments: mesos-agent-flamegraph.png, mesos-agent.stacks.gz
>
>
> /proc/mounts is read many, many times from src/(linux/fs|linux/cgroups|slave/slave).cpp.
> When using overlayfs, the /proc/mounts contents can become quite large. 
> As an example, one of our Q/A single node running ~150 tasks,  have a 361 lines/ 201299 chars  /proc/mounts file.
> This 200kB file is read on this node about 25 to 150 times per second. This is a (huge) waste of cpu and I/O time.
> Most of these calls are related to cgroups.
> Please consider these proposals :
> 1/ Is /proc/mounts mandatory for cgroups ? 
> We already have cgroup subsystems list from /proc/cgroups.
> The only compelling information from /proc/mounts seems to be the root mount point, 
> /sys/fs/cgroup/, which could be obtained by a unique read on agent start.
> 2/ use /proc/self/mountstats
> {noformat}
> wc /proc/self/mounts /proc/self/mountstats
> 361 2166 201299 /proc/self/mounts
> 361 2888 50200 /proc/self/mountstats
> {noformat}
> {noformat}
> grep cgroup /proc/self/mounts
> cgroup /sys/fs/cgroup tmpfs rw,relatime,mode=755 0 0
> cgroup /sys/fs/cgroup/cpuset cgroup rw,relatime,cpuset 0 0
> cgroup /sys/fs/cgroup/cpu cgroup rw,relatime,cpu 0 0
> cgroup /sys/fs/cgroup/cpuacct cgroup rw,relatime,cpuacct 0 0
> cgroup /sys/fs/cgroup/blkio cgroup rw,relatime,blkio 0 0
> cgroup /sys/fs/cgroup/memory cgroup rw,relatime,memory 0 0
> cgroup /sys/fs/cgroup/devices cgroup rw,relatime,devices 0 0
> cgroup /sys/fs/cgroup/freezer cgroup rw,relatime,freezer 0 0
> cgroup /sys/fs/cgroup/net_cls cgroup rw,relatime,net_cls 0 0
> cgroup /sys/fs/cgroup/perf_event cgroup rw,relatime,perf_event 0 0
> cgroup /sys/fs/cgroup/net_prio cgroup rw,relatime,net_prio 0 0
> cgroup /sys/fs/cgroup/pids cgroup rw,relatime,pids 0 0
> {noformat}
> {noformat}
> grep cgroup /proc/self/mountstats
> device cgroup mounted on /sys/fs/cgroup with fstype tmpfs
> device cgroup mounted on /sys/fs/cgroup/cpuset with fstype cgroup
> device cgroup mounted on /sys/fs/cgroup/cpu with fstype cgroup
> device cgroup mounted on /sys/fs/cgroup/cpuacct with fstype cgroup
> device cgroup mounted on /sys/fs/cgroup/blkio with fstype cgroup
> device cgroup mounted on /sys/fs/cgroup/memory with fstype cgroup
> device cgroup mounted on /sys/fs/cgroup/devices with fstype cgroup
> device cgroup mounted on /sys/fs/cgroup/freezer with fstype cgroup
> device cgroup mounted on /sys/fs/cgroup/net_cls with fstype cgroup
> device cgroup mounted on /sys/fs/cgroup/perf_event with fstype cgroup
> device cgroup mounted on /sys/fs/cgroup/net_prio with fstype cgroup
> device cgroup mounted on /sys/fs/cgroup/pids with fstype cgroup
> {noformat}
> This file contains all the required information, and is 4x smaller
> 3/ microcaching
> Caching cgroups data for just 1 second would be a huge perfomance improvement, but i'm not aware of the possible side effects.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)