[
https://issues.apache.org/jira/browse/HADOOP-11873?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16182786#comment-16182786
]
Steve Loughran commented on HADOOP-11873:
-----------------------------------------
HADOOP-13065 has obsoleted this: we have an extensible map for statistic
collection. That could be collected from committers &aggregated.
That doesn't cover timing, and you don't necessarily just want the total or ave
numbers, it's that distribution which can show problems.
What could be done here is probably to use HTrace all the way down the chain.
HDFS already supports it both client-side and server side, the object stores
could add it client-side. All they need is the query engines (MR, Tez, Spark,
to propagate an HTrace span from the driver to the workers with the actual
task...it's only a tuple of longs, so trivial to include.
> Include disk read/write time in FileSystem.Statistics
> -----------------------------------------------------
>
> Key: HADOOP-11873
> URL: https://issues.apache.org/jira/browse/HADOOP-11873
> Project: Hadoop Common
> Issue Type: New Feature
> Components: metrics
> Reporter: Kay Ousterhout
> Priority: Minor
>
> Measuring the time spent blocking on reading / writing data from / to disk is
> very useful for debugging performance problems in applications that read data
> from Hadoop, and can give much more information (e.g., to reflect disk
> contention) than just knowing the total amount of data read. I'd like to add
> something like "diskMillis" to FileSystem#Statistics to track this.
> For data read from HDFS, this can be done with very low overhead by adding
> logging around calls to RemoteBlockReader2.readNextPacket (because this reads
> larger chunks of data, the time added by the instrumentation is very small
> relative to the time to actually read the data). For data written to HDFS,
> this can be done in DFSOutputStream.waitAndQueueCurrentPacket.
> As far as I know, if you want this information today, it is only currently
> accessible by turning on HTrace. It looks like HTrace can't be selectively
> enabled, so a user can't just turn on the tracing on
> RemoteBlockReader2.readNextPacket for example, and instead needs to turn on
> tracing everywhere (which then introduces a bunch of overhead -- so sampling
> is necessary). It would be hugely helpful to have native metrics for time
> reading / writing to disk that are sufficiently low-overhead to be always on.
> (Please correct me if I'm wrong here about what's possible today!)
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
