On Mon, 2004-08-09 at 19:23, Florin Andrei wrote: > On Mon, 2004-08-09 at 13:48, Florian Schmidt wrote: > > > Hmm, in my case the maximum value i can use is 128, and i cannot see a > > difference between 128 and 16 either. It didn't do harm. > > Maybe it vastly depends on the workload? You know, like writing a tiny > 44.1/16/2 audio stream to disk versus writing a huge raw video stream > (or a large database workload). > > What was your disk workload? I tried every disk torture benchmark I could think of - bonnie++, sysbench, dbench, iozone. I also tried stressing the CPU/scheduler with parallel builds and CPU stress tests. Stressing the system increases the average latency slightly (I have graphs at http://www.members.dca.net/rlrevell/testresults/) but does not affect the worst case at all. I am guessing whatever code path is involved in the higher latencies is not filesystem-related. The 128 vs 1024 default has to do with whether 48-bit addressing is enabled. This will be enabled for larger disks (120GB+?). So, the size of your disk affects the maximum amount of IO that can be in flight at once. This was covered in one of the recent lkml threads, check the archives. Lee
