[
https://issues.apache.org/jira/browse/HDFS-8873?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14903507#comment-14903507
]
Colin Patrick McCabe commented on HDFS-8873:
--------------------------------------------
bq. I didn't do that in this patch because I didn't think the 1000 was as
prominent as before, but it appears that was before I was done adding stuff.
I'll put it back. I'd love to put that constant somewhere like util.Time. Would
that be kosher?
I think it's fine to put it in the DirectoryScanner itself if you want. I
don't object to putting it in time either. Up to you.
bq. I was shooting for something that would be meaningful to someone who
doesn't know the code. What about "throttle limit," since that echoes the
config param?
Sure.
bq. I don't follow. (nowMs % 1000) has to be between 0 and 999. If it's less
than the throttle limit, we won't enter the loop. The throttle limit must be
between 1 and 1000. (Anything else gets set to 1000 when the scanner is
created.) The sleep must therefore be for between 1 and 999 ms, pretty much
guaranteeing a different result the next time around.
Let's say we start the loop at time 5200. Then {{while (nowMs % 1000L >
throttleLimitMsPerSec)}} returns true (let's say {{throttleLimitMsPerSec =
100}}).
We call sleep with an argument of 800, but sleep actually sleeps for 1000 ms
instead. (Remember, Thread#sleep may always sleep for longer than requested.)
nowMs becomes 6200. Now {{while (nowMs % 1000L > throttleLimitMsPerSec) }}
returns true again, since 6200 % 1000 = 200 > 100. Now we sleep again for 800
ms yet again. We completely missed our timeslice, and there's no guarantee
that we'll pick up the next one either. That's the bug.
> 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
>
>
> 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)
