If not, maybe I just generate the same 1,000,000 files on each machine, then randomly delete 1/2 the files and stream them from the other machine as writing those files would all be in random locations again forcing a much worse measurement of MB/sec I would think.
Not sure I understand the question. But you could just scrub the data off a node and rebuild it. 

Note that streaming is throttled, and it will also generate compaction.

He has twenty 1T drives on each machine and I think he also tried with one 1T drive seeing the same performance which makes sense if writing sequentially
Are you using the 1.2 JBOB configuration?

Cheers

-----------------
Aaron Morton
Freelance Cassandra Consultant
New Zealand

@aaronmorton

On 1/04/2013, at 11:01 PM, "Hiller, Dean" <Dean.Hiller@nrel.gov> wrote:

(we plan on running similar performance tests on cassandra but wanted to understand the raw foot print first)..

Someone in ops was doing a test transferring 1T of data from one node to another.  I had a huge concern I emailed him that this could end up being a completely sequential write not testing random access speeds.  He has twenty 1T drives on each machine and I think he also tried with one 1T drive seeing the same performance which makes sense if writing sequentially.  Does anyone know of something that could generate a random access pattern such that we could time that?  Right now, he was measuring 253MB / second from the time it took and the 1T of data.  I would like to find the much worse case of course.

If not, maybe I just generate the same 1,000,000 files on each machine, then randomly delete 1/2 the files and stream them from the other machine as writing those files would all be in random locations again forcing a much worse measurement of MB/sec I would think.

Thoughts?

Thanks,
Dean