[
https://issues.apache.org/jira/browse/HDFS-8790?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15273002#comment-15273002
]
Xiaowei Zhu commented on HDFS-8790:
-----------------------------------
Per James,
1. Ideally it should stat the file it's running tests against so it knows the
upper bounds for seek rather than a hard coded max offset.
2. Lots of stuff can be cleaned up, and then also adding more test cases like
large reads.
3. And to emulate cancels in vertica have a whole bunch of threads doing reads
like EE would be doing, and then cancel them all at once.
4. GetBlockLocations has the file length so that could be used instead of stat
temporarily.
> Add Filesystem level stress tests
> ---------------------------------
>
> Key: HDFS-8790
> URL: https://issues.apache.org/jira/browse/HDFS-8790
> Project: Hadoop HDFS
> Issue Type: Sub-task
> Components: hdfs-client
> Reporter: James Clampffer
> Assignee: James Clampffer
> Attachments: HDFS-8790.HDFS-8707.000.patch
>
>
> I propose adding stress tests on libhdfs(3) compatibility layer was well as
> the async calls. These can be also used for basic performance metrics and
> inputs to profiling tools to see improvements over time.
> I'd like to make these tests into a seperate executable, or set of them, so
> that they can be used for longer running tests on dedicated clusters that may
> already exist. Each should provide a simple command line interface for
> scripted or manual use.
> Basic tests would be:
> looped open-read-close
> sequential scans
> small random reads
> All tests will be parameterized for number of threads, read size, and upper
> and lower offset bounds for a specified file. This will make it much easier
> to detect and reproduce threading issues and resource leaks as well as
> provide a simple executable (or set of executables) that can be run with
> valgrind to gain a high confidence that the code is operating correctly.
> I'd appreciate suggestions for any other simple stress tests.
> HDFS-8766 intentionally avoided shared_ptr and unique_ptr in the C api to
> make debugging this a little easier in case memory stomps and dangling
> references show up in stress tests. These will be added into the C API when
> the patch for this jira is submitted because things should be reasonably
> stable once the stress tests pass.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: hdfs-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: hdfs-issues-h...@hadoop.apache.org
