Jeremy Villalobos 2011-08-29, 05:07
-Re: Running scale tests
Robert Evans 2011-08-31, 15:31
That really depends on how accurate you want your tests to be vs. how long you want to spend running them. Often automated OS benchmarks will reinstall the OS before running tests to try to ensure it is in a known state. File system bench marks usually will unmount the file system and then remount it to ensure that caches are empty. If you want to be very accurate then reformat HDFS and reconfigure everything. However even in ideal situations it is difficult to get consistent performance numbers out of a multinode cluster. I would suggest you bring up the cluster with the maximum number of nodes you want to test with, then shut down some of the data nodes and task trackers on the machines you don't want (Just like if the box died). Then wait for the data on them to finish being replicated. It should be fairly close to what you would expect .
You can also do it programmatically. You can black list machines using the admin interface on the task tracker and name node, but I have not done it before.
On 8/29/11 12:07 AM, "Jeremy Villalobos" <[EMAIL PROTECTED]> wrote:
The following questions are from an system administrator point of view.
How do I run scale tests using different numbers of nodes ? Do I have to shutdown and restart hadoop to do this ?
What about dfs, do I have to reformat when changing the number of nodes down ?
Is there a "machines file" as done in MPI where I can specify just the number of nodes to be used for a test ?