On Thu, Jun 11, 2015 at 2:23 PM, Ilya Dryomov <idryomov@xxxxxxxxx> wrote: > On Wed, Jun 10, 2015 at 7:07 PM, Ilya Dryomov <idryomov@xxxxxxxxx> wrote: >> On Wed, Jun 10, 2015 at 7:04 PM, Nick Fisk <nick@xxxxxxxxxx> wrote: >>>> > >> -----Original Message----- >>>> > >> From: Ilya Dryomov [mailto:idryomov@xxxxxxxxx] >>>> > >> Sent: 10 June 2015 14:06 >>>> > >> To: Nick Fisk >>>> > >> Cc: ceph-users >>>> > >> Subject: Re: krbd splitting large IO's into smaller >>>> > >> IO's >>>> > >> >>>> > >> On Wed, Jun 10, 2015 at 2:47 PM, Nick Fisk <nick@xxxxxxxxxx> wrote: >>>> > >> > Hi, >>>> > >> > >>>> > >> > Using Kernel RBD client with Kernel 4.03 (I have also tried some >>>> > >> > older kernels with the same effect) and IO is being split into >>>> > >> > smaller IO's which is having a negative impact on performance. >>>> > >> > >>>> > >> > cat /sys/block/sdc/queue/max_hw_sectors_kb >>>> > >> > 4096 >>>> > >> > >>>> > >> > cat /sys/block/rbd0/queue/max_sectors_kb >>>> > >> > 4096 >>>> > >> > >>>> > >> > Using DD >>>> > >> > dd if=/dev/rbd0 of=/dev/null bs=4M >>>> > >> > >>>> > >> > Device: rrqm/s wrqm/s r/s w/s rkB/s wkB/s >>>> avgrq-sz >>>> > >> > avgqu-sz await r_await w_await svctm %util >>>> > >> > rbd0 0.00 0.00 201.50 0.00 25792.00 0.00 >>>> 256.00 >>>> > >> > 1.99 10.15 10.15 0.00 4.96 100.00 >>>> > >> > >>>> > >> > >>>> > >> > Using FIO with 4M blocks >>>> > >> > Device: rrqm/s wrqm/s r/s w/s rkB/s wkB/s >>>> avgrq-sz >>>> > >> > avgqu-sz await r_await w_await svctm %util >>>> > >> > rbd0 0.00 0.00 232.00 0.00 118784.00 0.00 >>>> 1024.00 >>>> > >> > 11.29 48.58 48.58 0.00 4.31 100.00 >>>> > >> > >>>> > >> > Any ideas why IO sizes are limited to 128k (256 blocks) in DD's >>>> > >> > case and 512k in Fio's case? >>>> > >> >>>> > >> 128k vs 512k is probably buffered vs direct IO - add iflag=direct >>>> > >> to your dd invocation. >>>> > > >>>> > > Yes, thanks for this, that was the case >>>> > > >>>> > >> >>>> > >> As for the 512k - I'm pretty sure it's a regression in our switch >>>> > >> to blk-mq. I tested it around 3.18-3.19 and saw steady 4M IOs. I >>>> > >> hope we are just missing a knob - I'll take a look. >>>> > > >>>> > > I've tested both 4.03 and 3.16 and both seem to be split into 512k. >>>> > > Let >>>> me >>>> > know if you need me to test any other particular version. >>>> > >>>> > With 3.16 you are going to need to adjust max_hw_sectors_kb / >>>> > max_sectors_kb as discussed in Dan's thread. The patch that fixed >>>> > that in the block layer went into 3.19, blk-mq into 4.0 - try 3.19. >>>> >>>> Sorry should have mentioned, I had adjusted both of them on the 3.16 >>>> kernel to 4096. >>>> I will try 3.19 and let you know. >>> >>> Better with 3.19, but should I not be seeing around 8192, or am I getting my >>> blocks and bytes mixed up? >>> >>> Device: rrqm/s wrqm/s r/s w/s rkB/s wkB/s avgrq-sz >>> avgqu-sz await r_await w_await svctm %util >>> rbd0 72.00 0.00 24.00 0.00 49152.00 0.00 4096.00 >>> 1.96 82.67 82.67 0.00 41.58 99.80 >> >> I'd expect 8192. I'm getting a box for investigation. > > OK, so this is bug in the blk-mq part of block layer. There is no > plugging going on in the single hardware queue (i.e. krbd) case - it > never once plugs the queue, and that means no request merging is done > for your direct sequential read test. It gets 512k bios and those same > 512k requests are issued to krbd. > > While queue plugging may not make sense in the multi queue case, I'm > pretty sure it's supposed to plug in the single queue case. Looks like > use_plug logic in blk_sq_make_request() is busted. It turns out to be a year old regression. Before commit 07068d5b8ed8 ("blk-mq: split make request handler for multi and single queue") it used to be (reads are considered sync) use_plug = !is_flush_fua && ((q->nr_hw_queues == 1) || !is_sync); and now it is use_plug = !is_flush_fua && !is_sync; in a function that is only called if q->nr_hw_queues == 1. This is getting fixed by "blk-mq: fix plugging in blk_sq_make_request" from Jeff Moyer - http://article.gmane.org/gmane.linux.kernel/1941750. Looks like it's on its way to mainline along with some other blk-mq plugging fixes. Thanks, Ilya _______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
 |