[
https://issues.apache.org/jira/browse/HADOOP-15901?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16675133#comment-16675133
]
Steve Loughran commented on HADOOP-15901:
-----------------------------------------
-1
I used to think MonotonicNow was good, but now I don't.
h3. Monotonic now is not guaranteed to be consistent across cores in a single
socket, absolutely not consistent across sockets
It is only safe to use in certain situations
* you are doing micro benchmarks where you can take averages of many runs
discard outliers
* you want a very low cost time value for diagnostics and testing & can cope
with inconsistency, and you are confident the probability of a thread being
suspended and rescheduled is 0
* you are running code where you have used CPU affinity calls to bind a thread
to a specific core, and you are confident that you are not running in a VM,
where the host OS can remap virtual to physical cores as it chooses.
Given IPC blocks for network requests, I'm not sure it matches any of these.
Further reading:
http://steveloughran.blogspot.com/2015/09/time-on-multi-core-multi-socket-servers.html
> IPC Client and Server should use Time.monotonicNow() for elapsed times.
> -----------------------------------------------------------------------
>
> Key: HADOOP-15901
> URL: https://issues.apache.org/jira/browse/HADOOP-15901
> Project: Hadoop Common
> Issue Type: Bug
> Components: ipc, metrics
> Reporter: Vinayakumar B
> Assignee: Vinayakumar B
> Priority: Major
> Attachments: HADOOP-15901-01.patch
>
>
> Client.java and Server.java uses {{Time.now()}} to calculate the elapsed
> times/timeouts. This could result in undesired results when system clock's
> time changes.
> {{Time.monotonicNow()}} should be used for elapsed time calculations within
> same JVM.
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
