| Message ID | 1503326134-3862-1-git-send-email-sbates@raithlin.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On 08/21/2017 08:35 AM, sbates@raithlin.com wrote: > From: Stephen Bates <sbates@raithlin.com> > > Hybrid polling currently uses half the average completion time as an > estimate of how long to poll for. We can improve upon this by noting > that polling before the minimum completion time makes no sense. Add a > sysfs entry to use this fact to improve CPU utilization in certain > cases. > > At the same time the minimum is a bit too long to sleep for since we > must factor in OS wake time for the thread. For now allow the user to > set this via a second sysfs entry (in nanoseconds). > > Testing this patch on Intel Optane SSDs showed that using the minimum > rather than half reduced CPU utilization from 59% to 38%. Tuning > this via the wake time adjustment allowed us to trade CPU load for > latency. For example > > io_poll delay hyb_use_min adjust latency CPU load > 1 -1 N/A N/A 8.4 100% > 1 0 0 N/A 8.4 57% > 1 0 1 0 10.3 34% > 1 9 1 1000 9.9 37% > 1 0 1 2000 8.4 47% > 1 0 1 10000 8.4 100% > > Ideally we will extend this to auto-calculate the wake time rather > than have it set by the user. I don't like this, it's another weird knob that will exist but that no one will know how to use. For most of the testing I've done recently, hybrid is a win over busy polling - hence I think we should make that the default. 60% of mean has also, in testing, been shown to be a win. So that's an easy fix/change we can consider. To go beyond that, I'd much rather see us tracking the time waste. If we consider the total completion time of an IO to be A+B+C, where: A Time needed to go to sleep B Sleep time C Time needed to wake up then we could feasibly track A+C. We already know how long the IO will take to complete, as we track that. At that point we'd have a full picture of how long we should sleep. Bonus points for informing the lower level scheduler of this as well. If the CPU is going idle, we'll enter some sort of power state in the processor. If we were able to pass in how long we expect to sleep, we could be making better decisions here.
>> From: Stephen Bates <sbates@raithlin.com> >> >> Hybrid polling currently uses half the average completion time as an >> estimate of how long to poll for. We can improve upon this by noting >> that polling before the minimum completion time makes no sense. Add a >> sysfs entry to use this fact to improve CPU utilization in certain >> cases. >> >> At the same time the minimum is a bit too long to sleep for since we >> must factor in OS wake time for the thread. For now allow the user to >> set this via a second sysfs entry (in nanoseconds). >> >> Testing this patch on Intel Optane SSDs showed that using the minimum >> rather than half reduced CPU utilization from 59% to 38%. Tuning >> this via the wake time adjustment allowed us to trade CPU load for >> latency. For example >> >> io_poll delay hyb_use_min adjust latency CPU load >> 1 -1 N/A N/A 8.4 100% >> 1 0 0 N/A 8.4 57% >> 1 0 1 0 10.3 34% >> 1 9 1 1000 9.9 37% >> 1 0 1 2000 8.4 47% >> 1 0 1 10000 8.4 100% >> >> Ideally we will extend this to auto-calculate the wake time rather >> than have it set by the user. > > I don't like this, it's another weird knob that will exist but that > no one will know how to use. For most of the testing I've done > recently, hybrid is a win over busy polling - hence I think we should > make that the default. 60% of mean has also, in testing, been shown > to be a win. So that's an easy fix/change we can consider. I do agree that the this is a hard knob to tune. I am however not happy that the current hybrid default may mean we are polling well before the minimum completion time. That just seems like a waste of CPU resources to me. I do agree that turning on hybrid as the default and perhaps bumping up the default is a good idea. > To go beyond that, I'd much rather see us tracking the time waste. > If we consider the total completion time of an IO to be A+B+C, where: > > A Time needed to go to sleep > B Sleep time > C Time needed to wake up > > then we could feasibly track A+C. We already know how long the IO > will take to complete, as we track that. At that point we'd have > a full picture of how long we should sleep. Yes, this is where I was thinking of taking this functionality in the long term. It seems like tracking C is something other parts of the kernel might need. Does anyone know of any existing code in this space? > Bonus points for informing the lower level scheduler of this as > well. If the CPU is going idle, we'll enter some sort of power > state in the processor. If we were able to pass in how long we > expect to sleep, we could be making better decisions here. Yup. Again, this seems like something more general that just the block-layer. I will do some digging and see/if anything is available to leverage here. Cheers Stephen
diff --git a/block/blk-mq.c b/block/blk-mq.c index f84d145..f453a35 100644 --- a/block/blk-mq.c +++ b/block/blk-mq.c @@ -2739,6 +2739,16 @@ static unsigned long blk_mq_poll_nsecs(struct request_queue *q, if (q->poll_stat[bucket].nr_samples) ret = (q->poll_stat[bucket].mean + 1) / 2; + if (q->poll_hyb_use_min) + ret = max(ret, (unsigned long)q->poll_stat[bucket].min); + + if (q->poll_hyb_adjust) { + if (ret >= q->poll_hyb_adjust) + ret -= q->poll_hyb_adjust; + else + return 0; + } + return ret; } diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c index 27aceab..51e5853 100644 --- a/block/blk-sysfs.c +++ b/block/blk-sysfs.c @@ -395,6 +395,50 @@ static ssize_t queue_poll_delay_store(struct request_queue *q, const char *page, return count; } +static ssize_t queue_poll_hyb_use_min_show(struct request_queue *q, char *page) +{ + return sprintf(page, "%d\n", q->poll_hyb_use_min); +} + +static ssize_t queue_poll_hyb_use_min_store(struct request_queue *q, + const char *page, size_t count) +{ + int err, val; + + if (!q->mq_ops || !q->mq_ops->poll) + return -EINVAL; + + err = kstrtoint(page, 10, &val); + if (err < 0) + return err; + + q->poll_hyb_use_min = val; + + return count; +} + +static ssize_t queue_poll_hyb_adjust_show(struct request_queue *q, char *page) +{ + return sprintf(page, "%d\n", q->poll_hyb_adjust); +} + +static ssize_t queue_poll_hyb_adjust_store(struct request_queue *q, + const char *page, size_t count) +{ + int err, val; + + if (!q->mq_ops || !q->mq_ops->poll) + return -EINVAL; + + err = kstrtoint(page, 10, &val); + if (err < 0) + return err; + + q->poll_hyb_adjust = val; + + return count; +} + static ssize_t queue_poll_show(struct request_queue *q, char *page) { return queue_var_show(test_bit(QUEUE_FLAG_POLL, &q->queue_flags), page); @@ -661,6 +705,18 @@ static ssize_t queue_dax_show(struct request_queue *q, char *page) .store = queue_poll_delay_store, }; +static struct queue_sysfs_entry queue_poll_hyb_use_min_entry = { + .attr = {.name = "io_poll_hyb_use_min", .mode = S_IRUGO | S_IWUSR }, + .show = queue_poll_hyb_use_min_show, + .store = queue_poll_hyb_use_min_store, +}; + +static struct queue_sysfs_entry queue_poll_hyb_adjust_entry = { + .attr = {.name = "io_poll_hyb_adjust", .mode = S_IRUGO | S_IWUSR }, + .show = queue_poll_hyb_adjust_show, + .store = queue_poll_hyb_adjust_store, +}; + static struct queue_sysfs_entry queue_wc_entry = { .attr = {.name = "write_cache", .mode = S_IRUGO | S_IWUSR }, .show = queue_wc_show, @@ -719,6 +775,8 @@ static ssize_t queue_dax_show(struct request_queue *q, char *page) &queue_dax_entry.attr, &queue_wb_lat_entry.attr, &queue_poll_delay_entry.attr, + &queue_poll_hyb_use_min_entry.attr, + &queue_poll_hyb_adjust_entry.attr, #ifdef CONFIG_BLK_DEV_THROTTLING_LOW &throtl_sample_time_entry.attr, #endif diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h index f45f157..97b46ce 100644 --- a/include/linux/blkdev.h +++ b/include/linux/blkdev.h @@ -527,6 +527,9 @@ struct request_queue { unsigned int rq_timeout; int poll_nsec; + int poll_hyb_use_min; + int poll_hyb_adjust; + struct blk_stat_callback *poll_cb; struct blk_rq_stat poll_stat[BLK_MQ_POLL_STATS_BKTS];