tez-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Siddharth Seth (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (TEZ-3754) With large ordered fetches, stalling shuffle could lead to slowdown of job
Date Tue, 11 Jul 2017 21:34:00 GMT

    [ https://issues.apache.org/jira/browse/TEZ-3754?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16083027#comment-16083027
] 

Siddharth Seth commented on TEZ-3754:
-------------------------------------

[~rajesh.balamohan] - what is the corner case in which this is hit? Each time there's a host
with pending input and a fetcher slot available?
I was somehow under the impression that we handled the WAIT case better. Wonder if it regressed
at some point.

Is the plan to track previous fetched inputs for size, so that the connection is not re-established?,
or maybe remember just the last fetched size for a host.

Does this block 0.9?

> With large ordered fetches, stalling shuffle could lead to slowdown of job
> --------------------------------------------------------------------------
>
>                 Key: TEZ-3754
>                 URL: https://issues.apache.org/jira/browse/TEZ-3754
>             Project: Apache Tez
>          Issue Type: Bug
>            Reporter: Rajesh Balamohan
>            Assignee: Rajesh Balamohan
>         Attachments: TEZ-3754.1.patch
>
>
> When large ordered fetches are getting downloaded, depending on the memory availability
in merger, it is possible to stall the shuffle ({{Type::WAIT}}). This happens after establishing
the connection,validation at server side etc. However, depending on memory availability allocator
could return {{Type::WAIT}} to retry fetching later. 
> In corner cases, retries happen very frequently and this creates pressure on the server
side. Server side populates the headers and starts sending the map output. This effort is
wasted in the client side as data is not downloaded. When multiple nodes have similar issue,
following connections timesout (180 seconds), which turns out to be
> very expensive.
> {noformat}
> 2017-06-06 04:27:49,574 [INFO] [fetcher {Map_1} #44] |orderedgrouped.FetcherOrderedGrouped|:
fetcher#44 - MergerManager returned Status.WAIT ...
> 2017-06-06 04:27:49,575 [INFO] [fetcher {Map_1} #31] |shuffle.HttpConnection|: for url=http://ctr-e133-1493418528701-76897-01-000005:13562/mapOutput?job=job_1496458948260_0050&reduce=1931&map=attempt_1496458948260_0050_1_00_000843_0_10007,attempt_1496458948260_0050_1_00_000841_1_10008
sent hash and receievd reply 0 ms
> 2017-06-06 04:27:49,575 [INFO] [fetcher {Map_1} #31] |orderedgrouped.FetcherOrderedGrouped|:
fetcher#31 - MergerManager returned Status.WAIT ...
> 2017-06-06 04:27:49,576 [INFO] [fetcher {Map_1} #43] |shuffle.HttpConnection|: for url=http://ctr-e133-1493418528701-76897-01-000005:13562/mapOutput?job=job_1496458948260_0050&reduce=1931&map=attempt_1496458948260_0050_1_00_000841_1_10008,attempt_1496458948260_0050_1_00_000843_0_10007
sent hash and receievd reply 0 ms
> 2017-06-06 04:27:49,576 [INFO] [fetcher {Map_1} #43] |orderedgrouped.FetcherOrderedGrouped|:
fetcher#43 - MergerManager returned Status.WAIT ...
> 2017-06-06 04:27:49,578 [INFO] [fetcher {Map_1} #32] |shuffle.HttpConnection|: for url=http://ctr-e133-1493418528701-76897-01-000005:13562/mapOutput?job=job_1496458948260_0050&reduce=1931&map=attempt_1496458948260_0050_1_00_000843_0_10007,attempt_1496458948260_0050_1_00_000841_1_10008
sent hash and receievd reply 0 ms
> 2017-06-06 04:27:49,578 [INFO] [fetcher {Map_1} #32] |orderedgrouped.FetcherOrderedGrouped|:
fetcher#32 - MergerManager returned Status.WAIT ...
> 2017-06-06 04:27:49,579 [INFO] [fetcher {Map_1} #42] |shuffle.HttpConnection|: for url=http://ctr-e133-1493418528701-76897-01-000005:13562/mapOutput?job=job_1496458948260_0050&reduce=1931&map=attempt_1496458948260_0050_1_00_000841_1_10008,attempt_1496458948260_0050_1_00_000843_0_10007
sent hash and receievd reply 0 ms
> 2017-06-06 04:27:49,579 [INFO] [fetcher {Map_1} #42] |orderedgrouped.FetcherOrderedGrouped|:
fetcher#42 - MergerManager returned Status.WAIT ...
> 2017-06-06 04:27:49,580 [INFO] [fetcher {Map_1} #33] |shuffle.HttpConnection|: for url=http://ctr-e133-1493418528701-76897-01-000005:13562/mapOutput?job=job_1496458948260_0050&reduce=1931&map=attempt_1496458948260_0050_1_00_000843_0_10007,attempt_1496458948260_0050_1_00_000841_1_10008
sent hash and receievd reply 0 ms
> 2017-06-06 04:27:49,580 [INFO] [fetcher {Map_1} #33] |orderedgrouped.FetcherOrderedGrouped|:
fetcher#33 - MergerManager returned Status.WAIT ...
> 2017-06-06 04:27:49,581 [INFO] [fetcher {Map_1} #41] |shuffle.HttpConnection|: for url=http://ctr-e133-1493418528701-76897-01-000005:13562/mapOutput?job=job_1496458948260_0050&reduce=1931&map=attempt_1496458948260_0050_1_00_000841_1_10008,attempt_1496458948260_0050_1_00_000843_0_10007
sent hash and receievd reply 0 ms
> 2017-06-06 04:27:49,581 [INFO] [fetcher {Map_1} #41] |orderedgrouped.FetcherOrderedGrouped|:
fetcher#41 - MergerManager returned Status.WAIT ...
> 2017-06-06 04:27:49,583 [INFO] [fetcher {Map_1} #34] |shuffle.HttpConnection|: for url=http://ctr-e133-1493418528701-76897-01-000005:13562/mapOutput?job=job_1496458948260_0050&reduce=1931&map=attempt_1496458948260_0050_1_00_000843_0_10007,attempt_1496458948260_0050_1_00_000841_1_10008
sent hash and receievd reply 0 ms
> 2017-06-06 04:27:49,583 [INFO] [fetcher {Map_1} #34] |orderedgrouped.FetcherOrderedGrouped|:
fetcher#34 - MergerManager returned Status.WAIT ...
> 2017-06-06 04:27:49,584 [INFO] [fetcher {Map_1} #40] |shuffle.HttpConnection|: for url=http://ctr-e133-1493418528701-76897-01-000005:13562/mapOutput?job=job_1496458948260_0050&reduce=1931&map=attempt_1496458948260_0050_1_00_000841_1_10008,attempt_1496458948260_0050_1_00_000843_0_10007
sent hash and receievd reply 0 ms
> 2017-06-06 04:27:49,584 [INFO] [fetcher {Map_1} #40] |orderedgrouped.FetcherOrderedGrouped|:
fetcher#40 - MergerManager returned Status.WAIT ...
> 2017-06-06 04:27:49,586 [INFO] [fetcher {Map_1} #35] |shuffle.HttpConnection|: for url=http://ctr-e133-1493418528701-76897-01-000005:13562/mapOutput?job=job_1496458948260_0050&reduce=1931&map=attempt_1496458948260_0050_1_00_000843_0_10007,attempt_1496458948260_0050_1_00_000841_1_10008
sent hash and receievd reply 0 ms
> 2017-06-06 04:27:49,586 [INFO] [fetcher {Map_1} #35] |orderedgrouped.FetcherOrderedGrouped|:
fetcher#35 - MergerManager returned Status.WAIT ...
> 2017-06-06 04:27:49,587 [INFO] [fetcher {Map_1} #39] |shuffle.HttpConnection|: for url=http://ctr-e133-1493418528701-76897-01-000005:13562/mapOutput?job=job_1496458948260_0050&reduce=1931&map=attempt_1496458948260_0050_1_00_000841_1_10008,attempt_1496458948260_0050_1_00_000843_0_10007
sent hash and receievd reply 0 ms
> 2017-06-06 04:27:49,587 [INFO] [fetcher {Map_1} #39] |orderedgrouped.FetcherOrderedGrouped|:
fetcher#39 - MergerManager returned Status.WAIT ...
> 2017-06-06 04:27:49,588 [INFO] [fetcher {Map_1} #36] |shuffle.HttpConnection|: for url=http://ctr-e133-1493418528701-76897-01-000005:13562/mapOutput?job=job_1496458948260_0050&reduce=1931&map=attempt_1496458948260_0050_1_00_000843_0_10007,attempt_1496458948260_0050_1_00_000841_1_10008
sent hash and receievd reply 0 ms
> 2017-06-06 04:27:49,588 [INFO] [fetcher {Map_1} #36] |orderedgrouped.FetcherOrderedGrouped|:
fetcher#36 - MergerManager returned Status.WAIT ...
> 2017-06-06 04:27:49,589 [INFO] [fetcher {Map_1} #38] |shuffle.HttpConnection|: for url=http://ctr-e133-1493418528701-76897-01-000005:13562/mapOutput?job=job_1496458948260_0050&reduce=1931&map=attempt_1496458948260_0050_1_00_000841_1_10008,attempt_1496458948260_0050_1_00_000843_0_10007
sent hash and receievd reply 0 ms
> 2017-06-06 04:27:49,589 [INFO] [fetcher {Map_1} #38] |orderedgrouped.FetcherOrderedGrouped|:
fetcher#38 - MergerManager returned Status.WAIT ...
> 2017-06-06 04:27:49,590 [INFO] [fetcher {Map_1} #37] |shuffle.HttpConnection|: for url=http://ctr-e133-1493418528701-76897-01-000005:13562/mapOutput?job=job_1496458948260_0050&reduce=1931&map=attempt_1496458948260_0050_1_00_000843_0_10007,attempt_1496458948260_0050_1_00_000841_1_10008
sent hash and receievd reply 0 ms
> {noformat}
> In such cases, it would be good to wait until {{usedMemory}} gets released during "next"
iteration as opposed to retrying immediately.



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)

Mime
View raw message