[
https://issues.apache.org/jira/browse/HDFS-14973?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16970673#comment-16970673
]
Konstantin Shvachko commented on HDFS-14973:
--------------------------------------------
Hey [~xkrogen] good analysis.
IIRC, the idea was to delay the first "wave" of getBlocks. The first 100 in
your example, which is the number dispatcher threads. Indeed the first 20 will
go right away without delay. This is how many calls we want to tolerate on the
NameNode at once. One second later another 20 getBlocks() will hit the
NameNode, and so on up to 100.
The next wave of dispatcher threads after 100 should not hit the NameNode right
away. It is supposed first to {{executePendingMove()}}, then call
{{getBlocks()}}. And {{executePendingMove()}} naturally throttles the
dispatcher, so it was not necessary to delay the subsequent ways. I remember it
worked. It is possible that {{executePendingMove()}} became faster due to
HDFS-11742, which I did not check, or something else had changed.
> Balancer getBlocks RPC dispersal does not function properly
> -----------------------------------------------------------
>
> Key: HDFS-14973
> URL: https://issues.apache.org/jira/browse/HDFS-14973
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: balancer & mover
> Affects Versions: 2.9.0, 2.7.4, 2.8.2, 3.0.0
> Reporter: Erik Krogen
> Assignee: Erik Krogen
> Priority: Major
> Attachments: HDFS-14973.000.patch, HDFS-14973.test.patch
>
>
> In HDFS-11384, a mechanism was added to make the {{getBlocks}} RPC calls
> issued by the Balancer/Mover more dispersed, to alleviate load on the
> NameNode, since {{getBlocks}} can be very expensive and the Balancer should
> not impact normal cluster operation.
> Unfortunately, this functionality does not function as expected, especially
> when the dispatcher thread count is low. The primary issue is that the delay
> is applied only to the first N threads that are submitted to the dispatcher's
> executor, where N is the size of the dispatcher's threadpool, but *not* to
> the first R threads, where R is the number of allowed {{getBlocks}} QPS
> (currently hardcoded to 20). For example, if the threadpool size is 100 (the
> default), threads 0-19 have no delay, 20-99 have increased levels of delay,
> and 100+ have no delay. As I understand it, the intent of the logic was that
> the delay applied to the first 100 threads would force the dispatcher
> executor's threads to all be consumed, thus blocking subsequent (non-delayed)
> threads until the delay period has expired. However, threads 0-19 can finish
> very quickly (their work can often be fulfilled in the time it takes to
> execute a single {{getBlocks}} RPC, on the order of tens of milliseconds),
> thus opening up 20 new slots in the executor, which are then consumed by
> non-delayed threads 100-119, and so on. So, although 80 threads have had a
> delay applied, the non-delay threads rush through in the 20 non-delay slots.
> This problem gets even worse when the dispatcher threadpool size is less than
> the max {{getBlocks}} QPS. For example, if the threadpool size is 10, _no
> threads ever have a delay applied_, and the feature is not enabled at all.
> This problem wasn't surfaced in the original JIRA because the test
> incorrectly measured the period across which {{getBlocks}} RPCs were
> distributed. The variables {{startGetBlocksTime}} and {{endGetBlocksTime}}
> were used to track the time over which the {{getBlocks}} calls were made.
> However, {{startGetBlocksTime}} was initialized at the time of creation of
> the {{FSNameystem}} spy, which is before the mock DataNodes are started. Even
> worse, the Balancer in this test takes 2 iterations to complete balancing the
> cluster, so the time period {{endGetBlocksTime - startGetBlocksTime}}
> actually represents:
> {code}
> (time to submit getBlocks RPCs) + (DataNode startup time) + (time for the
> Dispatcher to complete an iteration of moving blocks)
> {code}
> Thus, the RPC QPS reported by the test is much lower than the RPC QPS seen
> during the period of initial block fetching.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
