hadoop-mapreduce-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Milind Bhandarkar (Commented) (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (MAPREDUCE-3255) MR2 needs a counter analogous to "slot seconds"
Date Tue, 25 Oct 2011 18:06:32 GMT

    [ https://issues.apache.org/jira/browse/MAPREDUCE-3255?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13135295#comment-13135295

Milind Bhandarkar commented on MAPREDUCE-3255:

Container resource requirements are used by the RM to allocate them, and are logged in the
resource usage. Since the chargeback would be on the resource consumption as determined by
RM in allocation decisions, that is the unit to be exposed, rather than the actual usage of
that resource.

I consider hadoop metering similar to freight / post office model. There is a per unit volume
(or weight) charge, and a per distance traveled charge. Each of those have difference cost
models (i.e. not a straight line, but a curve). One needs to expose both of these as individual
metrics, so that the billing software can do the C1(v)*C2(d) multiplications, rather than
having to be restricted by a single C(v*d) function.
> MR2 needs a counter analogous to "slot seconds"
> -----------------------------------------------
>                 Key: MAPREDUCE-3255
>                 URL: https://issues.apache.org/jira/browse/MAPREDUCE-3255
>             Project: Hadoop Map/Reduce
>          Issue Type: Improvement
>          Components: mrv2
>    Affects Versions: 0.23.0
>            Reporter: Todd Lipcon
> In older versions of MR, we have a SLOT_MILLIS counter which counts how many seconds
each task (and job) used of map/reduce slots. Since we no longer have the concept of a "slot"
in MR2, we should change this counter to be something like "RAM-megabyte-seconds" - ie each
task is charged for its resource requirement * the number of seconds it occupied those resources.
> Ideally these would be collected by the NodeManager rather than the children, so that
they could be used for billing purposes, but a first cut of the counters through the MR framework
itself would probably still be useful in more trusting environments.

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


View raw message