[
https://issues.apache.org/jira/browse/CASSANDRA-234?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Sandeep Tata updated CASSANDRA-234:
-----------------------------------
Attachment: commons-math-1.1.jar
234.patch
Here's a tool that loosely mimics the workload of one of the apps we have
running.
1. The code is in Java should anyone choose to use a "heavy client" that knows
where to send a given request ... this typically cuts latency in half.
2. The workload consists of 2 phases --
The first is a load phase that does a whole bunch of batch inserts.
The second phase is a mix of inserts (new data and replacement of old data),
get_columns and get_slices. There is also the option to do deletes that is
turned off by default.
3. To keep things simple, only one column family is used, and no supercolumns
are written/read.
4. In the end, the total write bandwidth and a bunch of latency stats are
reported. I pulled in commons-math to get percentiles (median, 99th percentile
latency numbers).
5. All the configuration parameters are specified as in conf/workload.xml
I've found a simpler version of this this tool rather useful in the past in
locating bugs and now for testing performance.
> Performance regression tests
> ----------------------------
>
> Key: CASSANDRA-234
> URL: https://issues.apache.org/jira/browse/CASSANDRA-234
> Project: Cassandra
> Issue Type: New Feature
> Components: Tools
> Environment: all
> Reporter: Sandeep Tata
> Assignee: Sandeep Tata
> Attachments: 234.patch, commons-math-1.1.jar
>
>
> Since we have a few performance/efficiency tickets open
> (https://issues.apache.org/jira/browse/CASSANDRA-233,
> https://issues.apache.org/jira/browse/CASSANDRA-174), we should probably
> start using some basic benchmark for regression.
> VPork's an option.
> Johan has a description of how to get going with it:
> http://blog.oskarsson.nu/2009/05/vpork.html
> Jon Travis has a description at http://wiki.github.com/trav/vpork/vpork
> A performance graph that keeps getting better as we move forward is always
> fun :-)
> Can apache provide resources for a perf regression service? How do other
> projects set this up? Suggestions?
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
