[
https://issues.apache.org/jira/browse/YARN-3020?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14299184#comment-14299184
]
Peter D Kirchner commented on YARN-3020:
----------------------------------------
That expected usage you describe, and current implementation contains a basic
synchronization problem.
The client application's RPC updates requests to the RM before it receives the
containers newly assigned during that heartbeat.
Therefore, if (as is currently the case) the client calculates the total
requests, the total is too large by at least the number of matching incoming
assignments.
Per expected usage and current implementation, both add and remove cause this
obsolete, too-high total to be sent.
Cause or coincidence, I see applications (including but not limited to
distributedShell) making matching requests in a short interval and never
calling remove.
They receive the behavior they need, or closer to it, than the expected usage
would produce.
Further, in this API implementation/expected usage the remove API tries to
serve two purposes that are similar but not identical: to update the
client-side bookkeeping and to identify the request data to be sent to the
server. The problem here is that if there are only removes for allocated
containers, then the server-side bookkeeping is correct until the client sends
the total. The removes called for incoming assigned containers should not be
forwarded to the RM until there is at least one matching add, or a bona-fide
removal of a previously add-ed request.
I suppose the current implementation could be defended because its error is:
1) "only" too high by the number of matching incoming assignments,
2) persists "only" for the number of heartbeats it takes to clear the
out of sync condition
3) results in spurious allocations "only" once the application's
intentional matching requests were granted.
I maintain that spurious allocations are worst-case and especially damaging if
obtained by preemption.
I want to suggest an alternative that is simpler and accurate, and limited to
the AMRMClient and RM. The fact that the scheduler is updated by replacement
informs the choice of where Yarn should calculate that total for a matching
request.
The client is in a position to accurately calculate how much its current wants
differ from what it has asked for over its life.
This suggests a fix to the synchronization problem by having the client send
the net of add/remove requests it has accumulated over a heartbeat cycle,
and having the RM update its totals, from the difference obtained from the
client, using synchronized methods.
(Note, this client would not ordinarily call remove when it received a
container, as the scheduler has already
properly accounted for it when it made the allocation).
> n similar addContainerRequest()s produce n*(n+1)/2 containers
> -------------------------------------------------------------
>
> Key: YARN-3020
> URL: https://issues.apache.org/jira/browse/YARN-3020
> Project: Hadoop YARN
> Issue Type: Bug
> Components: client
> Affects Versions: 2.5.0, 2.6.0, 2.5.1, 2.5.2
> Reporter: Peter D Kirchner
> Original Estimate: 24h
> Remaining Estimate: 24h
>
> BUG: If the application master calls addContainerRequest() n times, but with
> the same priority, I get up to 1+2+3+...+n containers = n*(n+1)/2 . The most
> containers are requested when the interval between calls to
> addContainerRequest() exceeds the heartbeat interval of calls to allocate()
> (in AMRMClientImpl's run() method).
> If the application master calls addContainerRequest() n times, but with a
> unique priority each time, I get n containers (as I intended).
> Analysis:
> There is a logic problem in AMRMClientImpl.java.
> Although AMRMClientImpl.java, allocate() does an ask.clear() , on subsequent
> calls to addContainerRequest(), addResourceRequest() finds the previous
> matching remoteRequest and increments the container count rather than
> starting anew, and does an addResourceRequestToAsk() which defeats the
> ask.clear().
> From documentation and code comments, it was hard for me to discern the
> intended behavior of the API, but the inconsistency reported in this issue
> suggests one case or the other is implemented incorrectly.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
