[
https://issues.apache.org/jira/browse/HDFS-8873?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14904730#comment-14904730
]
Daniel Templeton commented on HDFS-8873:
----------------------------------------
Forgot to mention that this patch tries be fair about oversleeps when
throttling, but it doesn't do anything to credit for full seconds lost entirely
to oversleeps. For example, if the throttle is set to 100ms, the following
could happen:
.0000 - thread calls throttle(), no block
.0422 - thread calls throttle(), sleep for 588ms
.5016 - thread wakes up from oversleep, run limit this second set to 116ms
.5219 - thread calls throttle(), sleep for 781ms
In the second that the thread wakes up from the oversleep, we try to give it
its full run time, but we don't credit it for full seconds lost. Notice also
that threads only offer to block once per cycle, so setting a very low throttle
limit will just make the threads run one cycle between sleeps.
The following could also happen:
.0000 - thread calls throttle(), no block
.0422 - thread calls throttle(), sleep for 588ms
.1970 - thread wakes up from oversleep, run limit this second set to 1070ms
.2089 - thread calls throttle(), no block
.2101 - thread calls throttle(), sleep for 899ms
When the thread wakes up from the oversleep, we give it more than a second's
worth of run time because it's so close to the end of the second. (Any value
over 999 just means not to throttle the thread that second.) The next second
starts the count over, and at .2089, even though it's been running for over 100
consecutive ms, we don't block it, because it hasn't run for 100ms in this
second yet. When it checks in the next time, at .2101, we see it's over the
100ms limit and block it for the rest of the second.
> throttle directoryScanner
> -------------------------
>
> Key: HDFS-8873
> URL: https://issues.apache.org/jira/browse/HDFS-8873
> Project: Hadoop HDFS
> Issue Type: Improvement
> Components: datanode
> Affects Versions: 2.7.1
> Reporter: Nathan Roberts
> Assignee: Daniel Templeton
> Attachments: HDFS-8873.001.patch, HDFS-8873.002.patch,
> HDFS-8873.003.patch, HDFS-8873.004.patch, HDFS-8873.005.patch,
> HDFS-8873.006.patch
>
>
> The new 2-level directory layout can make directory scans expensive in terms
> of disk seeks (see HDFS-8791) for details.
> It would be good if the directoryScanner() had a configurable duty cycle that
> would reduce its impact on disk performance (much like the approach in
> HDFS-8617).
> Without such a throttle, disks can go 100% busy for many minutes at a time
> (assuming the common case of all inodes in cache but no directory blocks
> cached, 64K seeks are required for full directory listing which translates to
> 655 seconds)
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
