Hi, since Jens didn't merge the patches yet, I'm sending v2 of the series. Since v1 I have added Jeff's Reviewed-by tag and also fixed two 32-bit issues (possible overflow when computing cfq_latency_target and 64-bit division using '/' in choose_wl_class_and_type()). Jens, please merge these patches. Thanks! Honza --- Intro to the series This series contains patches which convert CFQ to internally use nanoseconds instead of jiffies (based on patch from Jeff), use high resultion timers, and provide interfaces to tune slice length in microseconds instead of miliseconds. Currently, since CFQ uses jiffies interally, the precision of slice length is in the order of microseconds (e.g., 4 ms for SUSE kernels). In some cases it would be desirable to tune the slice length to 1 ms or even somewhat lower - especially in cases where the idling is necessary only to avoid starving processes doing dependent IO. There giving process 1 ms to submit more IO is usually more than enough. I have been running some IO benchmarks (dbench, postmark, tiobench) with these patches and didn't see any negative effect. I have also checked that tuning slice lenght down to 0.5 ms basically removes negative performance effect of isolating dbench process in a blkio cgroup in my setup. The optimal length of timeslice obviously depends on the setup and workload and it is not goal of this patch set to argue about the time slice lenght. Mostly I just wanted to demonstrate that tuning the timeslice length down helps some workloads and we don't want to disable idling completely. -- To unsubscribe from this list: send the line "unsubscribe linux-block" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
