On Tue, 2011-04-26 at 22:44 -0400, Justin Sherrill wrote: > For testing - Dillon's previous disk work has used blogbench for > testing, and I've used sysbench. There's also the usual task of > buildworld, though I suppose the next question would be how you'd > measure how well it worked in terms of interactivity. >
I have read about how the performance BFQ scheduler on Linux is measured[1][2]. I think we can design the following test cases: A. fairness oriented tests: Measure average disk bandwidth distribution of several parallel processes/threads that read/write same files of same size (and in continuous blocks on a disk). Although it quite differs from the actual cases that we meet everyday, it still can help to tell how fair a Fair Queue is. B. tests based on actual cases: - We can select a set of applications to test their response latency when the computer is under heavy I/Os, or just measure the time consumed for a cold start of an application (avoid being affected by the disk-memory mapping cache) - We can measure the performance of some applications that frequently and heavily do I/Os, such as DBMS. [1]http://algo.ing.unimo.it/people/paolo/disk_sched/results.php [2]http://algo.ing.unimo.it/people/paolo/disk_sched/comparison.php Brills Peng
