cfq: enable idle for seeky processes on rotational NCQ devices
PATCH 1/1 cfq: enable idle for seeky processes on rotational NCQ devices
Disabled idle window cause large latencies when seeky processes are competing with async writes, for rotational NCQ devices.
This patch, based on Vivek Goyal's original idea (http://lkml.org/lkml/2009/7/12/110),
unconditionally enables idle window for seeky processes on rotational NCQ devices. As for non-NCQ devices, a smaller idle window (2ms) is used
for seeky processes compared to normal I/O (8ms).
RAIDs should be marked as non-rotational as well (and probably a better flag name should be devised), since they can carry multiple operations in parallel.
---
block/cfq-iosched.c +6 -2
@@ -1981,10 +1981,14 @@ cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq, enable_idle = old_idle =cfq_cfqq_idle_window (cfqq);if (!atomic_read (&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle || - (!cfqd->cfq_latency && cfqd->hw_tag &&CIC_SEEKY (cic))) + (blk_queue_nonrot (cfqd->queue) && cfqd->hw_tag &&CIC_SEEKY (cic))) enable_idle =0 ;else if (sample_valid (cic->ttime_samples)) { -if (cic->ttime_mean > cfqd->cfq_slice_idle) +unsigned idle_time = cfqd->cfq_slice_idle; +if (CIC_SEEKY (cic)) + idle_time =CFQ_MIN_TT ; + +if (cic->ttime_mean > idle_time) enable_idle =0 ;else enable_idle =1 ;
Quoted 10 lines · Corrado Zoccolo
Disabled idle window cause large latencies when seeky processes are competing with async writes, for rotational NCQ devices.
This patch, based on Vivek Goyal's original idea (http://lkml.org/lkml/2009/7/12/110),
unconditionally enables idle window for seeky processes on rotational NCQ devices. As for non-NCQ devices, a smaller idle window (2ms) is used
for seeky processes compared to normal I/O (8ms).RAIDs should be marked as non-rotational as well (and probably a better flag name should be devised), since they can carry multiple operations in parallel.
I think this one is a bit problematic. What I'd like seeky processes to do is enable 'idle unless other sync read comes in' for such cases, otherwise it will cost us a lot of performance on the seeky vs seeky cases because we don't get to take advantage of queuing. It would be perfectly fine to continue waiting if just async IO comes in, but if we have other seeky readers then they should get a turn.
I realize that this does skew potential priority issues.
Signature · 1 line
Jens Axboe
Quoted 4 lines · Jens Axboe
I think this one is a bit problematic. What I'd like seeky processes to do is enable 'idle unless other sync read comes in' for such cases, otherwise it will cost us a lot of performance on the seeky vs seeky cases because we don't get to take advantage of queuing.
Are we sure that queuing is beneficial in this workload, on non-raid rotational devices?
If the seeks are still quite local (e.g. when accessing a single file), given that seek time is proportional to seek length, idling should provide higher throughput.
Anyway, I'm working on an other patch that will group together all seeky queues and dispatch them without idling, and idle only on the last one, so if you prefer, this can be postponed until the other patch is ready.
Thanks,
Corrado
Quoted 9 lines
It would be
perfectly fine to continue waiting if just async IO comes in, but if we have other seeky readers then they should get a turn.I realize that this does skew potential priority issues.
--
Jens Axboe
Quoted 11 lines · Corrado Zoccolo
Quoted 4 lines · Jens Axboe
I think this one is a bit problematic. What I'd like seeky processes to do is enable 'idle unless other sync read comes in' for such cases, otherwise it will cost us a lot of performance on the seeky vs seeky cases because we don't get to take advantage of queuing.
Are we sure that queuing is beneficial in this workload, on non-raid rotational devices?
If the seeks are still quite local (e.g. when accessing a single file), given that seek time is proportional to seek length, idling should provide higher throughput.
Yes very sure, seek time is only approximately proportional to seek length. With queuing, you potentially can account for rotational delay, which is an equally big factor in IO latency. For small seeks, it's easy the dominating factor even.
Quoted 4 lines
Anyway, I'm working on an other patch that will group together all seeky queues and dispatch them without idling, and idle only on the last one, so if you prefer, this can be postponed until the other patch is ready.
I think so.
Signature · 1 line
Jens Axboe