[
https://issues.apache.org/jira/browse/MAPREDUCE-3205?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Todd Lipcon updated MAPREDUCE-3205:
-----------------------------------
Attachment: mr-3205.txt
Here's a patch implementing the above suggestion:
- {{yarn.nodemanager.resource.memory-gb}} now determines the total physical
memory usage for the NM
- {{yarn.nodemanager.reserved.memory-mb}} no longer exists - it used to be used
to determine the total physical memory for the NM, but now obsoleted by the
above.
- {{yarn.nodemanager.vmem.to.pmem.limit.ratio}} sets the overcommit ratio for
vmem vs pmem. This ratio is used both for the total usage across tasks, and for
individual tasks.
- resource allocations requested by tasks now allocate physical memory instead
of virtual, with the virtual limit being determined by the above ratio.
I ran a 20 node test last night with the following parameters:
- {{yarn.nodemanager.resource.memory-gb}} = 16
- {{yarn.nodemanager.vmem.to.pmem.limit.ratio}} set to default (2.1)
- {{mapred.child.java.opts}} = -Xmx800m
- {{mapreduce.map.memory.mb}} set to default (1024)
This resulted in a pmem limit of 1024m and vmem limit of 2150m. I ran a
terasort and all of the tasks completed successfully. The VIRTUAL_MEMORY_BYTES
counter shows that most of the tasks were using around 1.2G of vmem. The
PHYSICAL_MEMORY_BYTES counter spans from about 500M to 950M.
Suggestions for how to add an appropriate unit test would be appreciated. Also,
if there's a better way of threading the Configuration object into
ContainerImpl, please let me know.
> MR2 memory limits should be pmem, not vmem
> ------------------------------------------
>
> Key: MAPREDUCE-3205
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-3205
> Project: Hadoop Map/Reduce
> Issue Type: Improvement
> Components: mrv2, nodemanager
> Affects Versions: 0.23.0
> Reporter: Todd Lipcon
> Attachments: mr-3205.txt
>
>
> Currently, the memory resources requested for a container limit the amount of
> virtual memory used by the container. On my test clusters, at least, Java
> processes take up nearly twice as much vmem as pmem - a Java process running
> with -Xmx500m uses 935m of vmem and only about 560m of pmem.
> This will force admins to either under-utilize available physical memory, or
> oversubscribe it by configuring the available resources on a TT to be larger
> than the true amount of physical RAM.
> Instead, I would propose that the resource limit apply to pmem, and allow the
> admin to configure a "vmem overcommit ratio" which sets the vmem limit as a
> function of pmem limit.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
