On Fri, Jun 05, 2020 at 10:21:31PM +0800, Hou Tao wrote: > Hi Ming, > > On 2020/6/4 18:01, Ming Lei wrote: > > Hi Hou Tao, > > > > On Wed, Jun 03, 2020 at 03:39:31PM +0800, Hou Tao wrote: > >> When there are many free-bit waiters, current batch wakeup method will > >> wake up at most wake_batch processes when wake_batch bits are freed. > >> The perfect result is each process will get a free bit, however the > >> real result is that a waken-up process may being unable to get > >> a free bit and will call io_schedule() multiple times. That's because > >> other processes (e.g. wake-up before) in the same wake-up batch > >> may have already allocated multiple free bits. > >> > >> And the race leads to two problems. The first one is the unnecessary > >> context switch, because multiple processes are waken up and then > >> go to sleep afterwards. And the second one is the performance > >> degradation when there is spatial locality between requests from > >> one process (e.g. split IO for HDD), because one process can not > >> allocated requests continuously for the split IOs, and > >> the sequential IOs will be dispatched separatedly. > > > > I guess this way is a bit worse for HDD since sequential IO may be > > interrupted by other context. > Yes. > > >> > >> To fix the problem, we mimic the way how SQ handles this situation: > > > > Do you mean the SQ way is the congestion control code in __get_request()? > > If not, could you provide more background of SQ's way for this issue? > > Cause it isn't easy for me to associate your approach with SQ's code. > > > The congestion control is accomplished by both __get_request() and __freed_request(). > In __get_request(), the max available requests is nr_requests * 1.5 when Actually, SQ code classified requests into sync an async, and for each type: the max allowed requests is nr_requests * 1.5, and batching allocation is triggered if rl->count[is_sync]+1 >= q->nr_requests or waking up from blocking allocation. > there are multiple threads try to allocate requests, and in __free_requests() > it only start to wake up waiter when the busy requests is less than nr_requests, > so half of nr_request is free when the waiter is woken-up. The SQ's batching allocation usually allows one active process to complete one batch of requests and others are blocked. This way is really nice for sequential IO on HDD. I did observe some HDD's writeback performance drops a lot after SQ's batching allocation is killed: [1] https://lore.kernel.org/linux-scsi/Pine.LNX.4.44L0.1909181213141.1507-100000@xxxxxxxxxxxxxxxxxxxx/ [2] https://lore.kernel.org/linux-scsi/20191226083706.GA17974@ming.t460p/ > > The approach in the patch is buggy, because it doesn't check whether > the number of busy bits is greater than the number of to-be-stashed > bits. So we just add an atomic (bit_busy) in struct sbitmap to track > the number of busy bits and use the number to decide whether Tracking busy bits is really expensive for SSD/NVMe, but it should be fine for HDD. Maybe we can one dedicated approach for HDD's request allocation. > we should wake one process or not: > > +#define SBQ_WS_ACTIVE_MIN 4 > + > +/* return true when fallback to batched wake-up is needed */ > +static bool sbitmap_do_stash_and_wakeup(struct sbitmap_queue *sbq) > +{ > + bool fall_back = false; > + int ws_active; > + struct sbq_wait_state *ws; > + int max_busy; > + int bit_busy; > + int wake_seq; > + int old; > + > + ws_active = atomic_read(&sbq->ws_active); > + if (!ws_active) > + goto done; > + > + if (ws_active < SBQ_WS_ACTIVE_MIN) { > + fall_back = true; > + goto done; > + } > + > + /* stash and make sure free bits >= depth / 4 */ > + max_busy = max_t(int, sbq->sb.depth * 3 / 4, 1); > + bit_busy = atomic_read(&sbq->bit_busy); > + if (bit_busy > max_busy) > + goto done; > + > +retry: > + ws = sbq_wake_ptr(sbq); > + if (!ws) > + goto done; > + > + wake_seq = atomic_read(&ws->wake_seq); > + old = atomic_cmpxchg(&ws->wake_seq, wake_seq, wake_seq + 1); > + if (old == wake_seq) { > + sbq_index_atomic_inc(&sbq->wake_index); > + wake_up(&ws->wait); > + goto done; > + } > + > +done: > + return fall_back; > +} > + > static bool __sbq_wake_up(struct sbitmap_queue *sbq) > { > struct sbq_wait_state *ws; > unsigned int wake_batch; > int wait_cnt; > > + if (sbq->flags & SBQ_FLAG_BATCH_BIT_ALLOC) { > + if (!sbitmap_do_stash_and_wakeup(sbq)) > + return false; > + } > + I feel that it is a good direction to add one such flag only for HDD's request tag allocation. > ws = sbq_wake_ptr(sbq); > if (!ws) > return false; > > >> 1) stash a bulk of free bits > >> 2) wake up a process when a new bit is freed > >> 3) woken-up process consumes the stashed free bits > >> 4) when stashed free bits are exhausted, goto step 1) > >>>> Because the tag allocation path or io submit path is much faster than > >> the tag free path, so when the race for free tags is intensive, > > > > Indeed, I guess you mean bio_endio is slow. > > > Yes, thanks for the correction. > > >> we can ensure: > >> 1) only few processes will be waken up and will exhaust the stashed > >> free bits quickly. > >> 2) these processes will be able to allocate multiple requests > >> continuously. > >> > >> An alternative fix is to dynamically adjust the number of woken-up > >> process according to the number of waiters and busy bits, instead of > >> using wake_batch each time in __sbq_wake_up(). However it will need > >> to record the number of busy bits all the time, so use the > >> stash-wake-use method instead. > >> > >> The following is the result of a simple fio test: > >> > >> 1. fio (random read, 1MB, libaio, iodepth=1024) > >> > >> (1) 4TB HDD (max_sectors_kb=256) > >> > >> IOPS (bs=1MB) > >> jobs | 4.18-sq | 5.6.15 | 5.6.15-patched | > >> 1 | 120 | 120 | 119 > >> 24 | 120 | 105 | 121 > >> 48 | 122 | 102 | 121 > >> 72 | 120 | 100 | 119 > >> > >> context switch per second > >> jobs | 4.18-sq | 5.6.15 | 5.6.15-patched | > >> 1 | 1058 | 1162 | 1188 > >> 24 | 1047 | 1715 | 1105 > >> 48 | 1109 | 1967 | 1105 > >> 72 | 1084 | 1908 | 1106 > >> > >> (2) 1.8TB SSD (set max_sectors_kb=256) > >> > >> IOPS (bs=1MB) > >> jobs | 4.18-sq | 5.6.15 | 5.6.15-patched | > >> 1 | 1077 | 1075 | 1076 > >> 24 | 1079 | 1075 | 1076 > >> 48 | 1077 | 1076 | 1076 > >> 72 | 1077 | 1076 | 1077 > >> > >> context switch per second > >> jobs | 4.18-sq | 5.6.15 | 5.6.15-patched | > >> 1 | 1833 | 5123 | 5264 > >> 24 | 2143 | 15238 | 3859 > >> 48 | 2182 | 19015 | 3617 > >> 72 | 2268 | 19050 | 3662 > >> > >> (3) 1.5TB nvme (set max_sectors_kb=256) > >> > >> 4 read queue, 72 CPU > >> > >> IOPS (bs=1MB) > >> jobs | 5.6.15 | 5.6.15-patched | > >> 1 | 3018 | 3018 > >> 18 | 3015 | 3016 > >> 36 | 3001 | 3005 > >> 54 | 2993 | 2997 > >> 72 | 2984 | 2990 > >> > >> context switch per second > >> jobs | 5.6.15 | 5.6.15-patched | > >> 1 | 6292 | 6469 > >> 18 | 19428 | 4253 > >> 36 | 21290 | 3928 > >> 54 | 23060 | 3957 > >> 72 | 24221 | 4054 > >> > >> Signed-off-by: Hou Tao <houtao1@xxxxxxxxxx> > >> --- > >> Hi, > >> > >> We found the problems (excessive context switch and few performance > >> degradation) during the performance comparison between blk-sq (4.18) > >> and blk-mq (5.16) on HDD, but we can not find a better way to fix it. > >> > >> It seems that in order to implement batched request allocation for > >> single process, we need to use an atomic variable to track > >> the number of busy bits. It's suitable for HDD or SDD, because the > >> IO latency is greater than 1ms, but no sure whether or not it's OK > >> for NVMe device. > > > > Do you have benchmark on NVMe/SSD with 4k BS? > > > The following is the randread test on SSD and NVMe. > > 1. fio randread 4KB > > (1) SSD 1.8TB (nr_tags=1024, nr_requests=256) > > It seems that when there is no race for tag allocation, the performance is the same, > but when there are intensive race for tag allocation, the performance gain is huge. > > total iodepth=256, so when jobs=2, iodepth=256/2=128 > > jobs | 5.6 | 5.6 patched > 1 | 193k | 192k > 2 | 197k | 196k > 4 | 198k | 198k > 8 | 197k | 197k > 16 | 197k | 198k > 32 | 198k | 198k > 64 | 195k | 195k > 128 | 193k | 192k > 256 | 198k | 198k > > total iodepth=512 > > jobs | 5.6 | 5.6 patched > 1 | 193k | 194k > 2 | 197k | 196k > 4 | 198k | 197k > 8 | 197k | 219k > 16 | 197k | 394k > 32 | 198k | 395k > 64 | 196k | 592k > 128 | 199k | 591k > 256 | 196k | 591k > 512 | 198k | 591k > > total iodepth=1024 > > jobs | 5.6 | 5.6 patched > 1 | 195k | 192k > 2 | 196k | 197k > 4 | 197k | 197k > 8 | 198k | 197k > 16 | 197k | 198k > 32 | 197k | 243k > 64 | 197k | 393k > 128 | 197k | 986k > 256 | 200k | 976k > 512 | 203k | 984k > 1024 | 202k | 354k > > (2) NVMe 1.5TB (nr_tags=1023) > > It seems there is no performance impact on NVMe device, but the > the number of context switch will be reduced. > > total iodepth=256, so when jobs=2, iodepth=256/2=128 > > jobs | 5.6 | 5.6 patched > 1 | 398k | 394k > 4 | 774k | 775k > 16 | 774k | 774k > 64 | 774k | 775k > 256 | 778k | 784k > > total iodepth=1024 > > jobs | 5.6 | 5.6 patched > 1 | 406k | 405k > 4 | 774k | 773k > 16 | 774k | 774k > 64 | 777k | 773k > 256 | 783k | 783k > 1024 | 764k | 755k > > total iodepth=2048 > > jobs | 5.6 | 5.6 patched > 1 | 369k | 377k > 4 | 774k | 774k > 16 | 774k | 774k > 64 | 767k | 773k > 256 | 784k | 781k > 1024 | 741k | 1416k > 2048 | 754k | 753k Frankly speaking, I am more interested in context switch & cpu utilization change on SSD/NVMe after applying your patch. We may improve HDD, meantime SSD/NVMe's perf can't be hurt, either latency, or cpu utilization. Thanks, Ming
