[
https://issues.apache.org/jira/browse/HDFS-10742?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15447450#comment-15447450
]
Chris Douglas commented on HDFS-10742:
--------------------------------------
bq. That adds a new object allocation for every lock acquire so it's probably
not very GC-friendly.
How so? If it's not inlined, the allocation is likely thread-local. Even if
neither is true, the allocation should be GC-friendly by being short-lived (the
critical section should be short). I could make up some nonsense about the
dispatch overhead for each call, or the cost of loading the class, but in the
context of a patch generating thread stacks and maintaining persistent maps of
timestamps, it's absurd to debate any of this.
> Measurement of lock held time in FsDatasetImpl
> ----------------------------------------------
>
> Key: HDFS-10742
> URL: https://issues.apache.org/jira/browse/HDFS-10742
> Project: Hadoop HDFS
> Issue Type: Improvement
> Components: datanode
> Affects Versions: 3.0.0-alpha2
> Reporter: Chen Liang
> Assignee: Chen Liang
> Attachments: HDFS-10742.001.patch, HDFS-10742.002.patch,
> HDFS-10742.003.patch, HDFS-10742.004.patch, HDFS-10742.005.patch,
> HDFS-10742.006.patch, HDFS-10742.007.patch, HDFS-10742.008.patch
>
>
> This JIRA proposes to measure the time the of lock of {{FsDatasetImpl}} is
> held by a thread. Doing so will allow us to measure lock statistics.
> This can be done by extending the {{AutoCloseableLock}} lock object in
> {{FsDatasetImpl}}. In the future we can also consider replacing the lock with
> a read-write lock.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
