* cfq-iosched.c:Use cfqq->nr_sectors in charge the vdisktime
@ 2011-03-30 15:23 Lina Lu
2011-03-30 15:54 ` Vivek Goyal
2011-03-31 15:46 ` Lina Lu
0 siblings, 2 replies; 6+ messages in thread
From: Lina Lu @ 2011-03-30 15:23 UTC (permalink / raw)
To: Vivek Goyal; +Cc: linux kernel mailing list
Hi Vivek,
I find the weight policy can be more accuracy with cfqq->nr_sectors instead
of cfqq->slice_dispatch.
Today, I try to modify cfq_group_served(), and use "charge = cfqq->nr_sectors; "
instead of "charge = cfqq->slice_dispatch; " . The test result seens more accuracy.
Why you choose slice_dispatch here? Is the nr_sectors will lower the total performance?
And in iops mod, if I try to apply weight policy on two IO processes with different
avgrq-sz, the test results will not exact match the weight value.
Thanks
Lina
^ permalink raw reply [flat|nested] 6+ messages in thread* Re: cfq-iosched.c:Use cfqq->nr_sectors in charge the vdisktime 2011-03-30 15:23 cfq-iosched.c:Use cfqq->nr_sectors in charge the vdisktime Lina Lu @ 2011-03-30 15:54 ` Vivek Goyal 2011-03-31 15:46 ` Lina Lu 1 sibling, 0 replies; 6+ messages in thread From: Vivek Goyal @ 2011-03-30 15:54 UTC (permalink / raw) To: Lina Lu; +Cc: linux kernel mailing list On Wed, Mar 30, 2011 at 11:23:30PM +0800, Lina Lu wrote: > Hi Vivek, > I find the weight policy can be more accuracy with cfqq->nr_sectors instead > of cfqq->slice_dispatch. > Today, I try to modify cfq_group_served(), and use "charge = cfqq->nr_sectors; " > instead of "charge = cfqq->slice_dispatch; " . The test result seens more accuracy. > Why you choose slice_dispatch here? Is the nr_sectors will lower the total performance? Lina, CFQ fundamentally allocates time slices hence accounting is done in time and not in terms of sectors. The other reason is that accounting in terms of time can be more accurate where some process is seeking all over the disk and doing little IO. If we account in terms of sectors then such seeky process will get much more share. > And in iops mod, if I try to apply weight policy on two IO processes with different > avgrq-sz, the test results will not exact match the weight value. IOPS mode kicks in when slice_idle=0. I suspect that group does not drive enough IO to remain on service tree hence gets deleted and hence loses share. Can you run a 20 sec backtrace and upload it somewhere. Thanks Vivek ^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: Re: cfq-iosched.c:Use cfqq->nr_sectors in charge the vdisktime 2011-03-30 15:23 cfq-iosched.c:Use cfqq->nr_sectors in charge the vdisktime Lina Lu 2011-03-30 15:54 ` Vivek Goyal @ 2011-03-31 15:46 ` Lina Lu 2011-03-31 19:46 ` Vivek Goyal 2011-04-01 14:59 ` Lina Lu 1 sibling, 2 replies; 6+ messages in thread From: Lina Lu @ 2011-03-31 15:46 UTC (permalink / raw) To: Vivek Goyal; +Cc: linux kernel mailing list On 2011-03-30 23:54:34, Vivek Goyal wrote: > On Wed, Mar 30, 2011 at 11:23:30PM +0800, Lina Lu wrote: > > Hi Vivek, > > I find the weight policy can be more accuracy with cfqq->nr_sectors instead > > of cfqq->slice_dispatch. > > Today, I try to modify cfq_group_served(), and use "charge = cfqq->nr_sectors; " > > instead of "charge = cfqq->slice_dispatch; " . The test result seens more accuracy. > > Why you choose slice_dispatch here? Is the nr_sectors will lower the total performance? > > Lina, > > CFQ fundamentally allocates time slices hence accounting is done in time > and not in terms of sectors. The other reason is that accounting in > terms of time can be more accurate where some process is seeking all > over the disk and doing little IO. If we account in terms of sectors > then such seeky process will get much more share. > > > And in iops mod, if I try to apply weight policy on two IO processes with different > > avgrq-sz, the test results will not exact match the weight value. > > IOPS mode kicks in when slice_idle=0. I suspect that group does not drive > enough IO to remain on service tree hence gets deleted and hence loses > share. > > Can you run a 20 sec backtrace and upload it somewhere. > Here is 20 sec backtrace: http://www.fileden.com/files/2010/9/9/2965145/cfq_log.tar.gz This time, I set two IO pid with weight 100, and the device is in iops_mod. linux-kzr4:/home/blkio # cat tst1/blkio.weight 100 linux-kzr4:/home/blkio # cat tst2/blkio.weight 100 iostat: Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util dm-0 0.00 0.00 855.50 0.00 3.34 0.00 8.00 0.82 1.06 0.95 81.70 dm-1 0.00 0.00 844.00 0.00 26.38 0.00 64.00 0.83 0.98 0.98 82.60 Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util dm-0 0.00 0.00 840.00 0.00 3.28 0.00 8.00 0.90 0.95 1.07 89.55 dm-1 0.00 0.00 794.00 0.00 24.81 0.00 64.00 0.87 1.10 1.10 87.00 Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util dm-0 0.00 0.00 596.50 0.00 2.33 0.00 8.00 0.96 1.77 1.61 95.80 dm-1 0.00 0.00 626.00 0.00 19.56 0.00 64.00 0.94 1.48 1.50 93.70 Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util dm-0 0.00 0.00 815.50 0.00 3.19 0.00 8.00 0.81 0.83 1.00 81.40 dm-1 0.00 0.00 828.50 0.00 25.89 0.00 64.00 0.77 0.95 0.93 77.45 Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util dm-0 0.00 0.00 910.50 0.00 3.56 0.00 8.00 0.82 1.00 0.90 82.15 dm-1 0.00 0.00 845.00 0.00 26.41 0.00 64.00 0.81 0.96 0.96 80.95 Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util dm-0 0.00 0.00 928.86 0.00 3.63 0.00 8.00 0.79 0.90 0.86 79.45 dm-1 0.00 0.00 848.26 0.00 26.51 0.00 64.00 0.65 0.77 0.77 65.17 >From the result, we can see that the iops match the weight value very well, but the rMB/s are not the same as they has different avgrq-sz. If I use the following patch, the rMB/s will be more accuracy. --- block/cfq-iosched.c 2011-03-31 23:43:55.000000000 +0800 +++ block/cfq-iosched.c 2011-03-31 23:44:30.000000000 +0800 @@ -951,7 +951,7 @@ used_sl = charge = cfq_cfqq_slice_usage(cfqq); if (iops_mode(cfqd)) - charge = cfqq->slice_dispatch; + charge = cfqq->nr_sectors; else if (!cfq_cfqq_sync(cfqq) && !nr_sync) charge = cfqq->allocated_slice; Thanks Lina ^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: Re: cfq-iosched.c:Use cfqq->nr_sectors in charge the vdisktime 2011-03-31 15:46 ` Lina Lu @ 2011-03-31 19:46 ` Vivek Goyal 2011-04-01 14:59 ` Lina Lu 1 sibling, 0 replies; 6+ messages in thread From: Vivek Goyal @ 2011-03-31 19:46 UTC (permalink / raw) To: Lina Lu; +Cc: linux kernel mailing list On Thu, Mar 31, 2011 at 11:46:37PM +0800, Lina Lu wrote: > On 2011-03-30 23:54:34, Vivek Goyal wrote: > > On Wed, Mar 30, 2011 at 11:23:30PM +0800, Lina Lu wrote: > > > Hi Vivek, > > > I find the weight policy can be more accuracy with cfqq->nr_sectors instead > > > of cfqq->slice_dispatch. > > > Today, I try to modify cfq_group_served(), and use "charge = cfqq->nr_sectors; " > > > instead of "charge = cfqq->slice_dispatch; " . The test result seens more accuracy. > > > Why you choose slice_dispatch here? Is the nr_sectors will lower the total performance? > > > > Lina, > > > > CFQ fundamentally allocates time slices hence accounting is done in time > > and not in terms of sectors. The other reason is that accounting in > > terms of time can be more accurate where some process is seeking all > > over the disk and doing little IO. If we account in terms of sectors > > then such seeky process will get much more share. > > > > > And in iops mod, if I try to apply weight policy on two IO processes with different > > > avgrq-sz, the test results will not exact match the weight value. > > > > IOPS mode kicks in when slice_idle=0. I suspect that group does not drive > > enough IO to remain on service tree hence gets deleted and hence loses > > share. > > > > Can you run a 20 sec backtrace and upload it somewhere. > > > > Here is 20 sec backtrace: > http://www.fileden.com/files/2010/9/9/2965145/cfq_log.tar.gz > > This time, I set two IO pid with weight 100, and the device is in iops_mod. How did you put device in iops mode? What's the device you are using and what kind of configuration dm-0 and dm-1 are in. > linux-kzr4:/home/blkio # cat tst1/blkio.weight > 100 > linux-kzr4:/home/blkio # cat tst2/blkio.weight > 100 > > iostat: > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > dm-0 0.00 0.00 855.50 0.00 3.34 0.00 8.00 0.82 1.06 0.95 81.70 > dm-1 0.00 0.00 844.00 0.00 26.38 0.00 64.00 0.83 0.98 0.98 82.60 > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > dm-0 0.00 0.00 840.00 0.00 3.28 0.00 8.00 0.90 0.95 1.07 89.55 > dm-1 0.00 0.00 794.00 0.00 24.81 0.00 64.00 0.87 1.10 1.10 87.00 > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > dm-0 0.00 0.00 596.50 0.00 2.33 0.00 8.00 0.96 1.77 1.61 95.80 > dm-1 0.00 0.00 626.00 0.00 19.56 0.00 64.00 0.94 1.48 1.50 93.70 > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > dm-0 0.00 0.00 815.50 0.00 3.19 0.00 8.00 0.81 0.83 1.00 81.40 > dm-1 0.00 0.00 828.50 0.00 25.89 0.00 64.00 0.77 0.95 0.93 77.45 > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > dm-0 0.00 0.00 910.50 0.00 3.56 0.00 8.00 0.82 1.00 0.90 82.15 > dm-1 0.00 0.00 845.00 0.00 26.41 0.00 64.00 0.81 0.96 0.96 80.95 > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > dm-0 0.00 0.00 928.86 0.00 3.63 0.00 8.00 0.79 0.90 0.86 79.45 > dm-1 0.00 0.00 848.26 0.00 26.51 0.00 64.00 0.65 0.77 0.77 65.17 > > >From the result, we can see that the iops match the weight value very well, but > the rMB/s are not the same as they has different avgrq-sz. > > If I use the following patch, the rMB/s will be more accuracy. > > --- block/cfq-iosched.c 2011-03-31 23:43:55.000000000 +0800 > +++ block/cfq-iosched.c 2011-03-31 23:44:30.000000000 +0800 > @@ -951,7 +951,7 @@ > used_sl = charge = cfq_cfqq_slice_usage(cfqq); > > if (iops_mode(cfqd)) > - charge = cfqq->slice_dispatch; > + charge = cfqq->nr_sectors; In IOPS mode we calculate the number of IOPS (that is number of requests dispatched) and not number of sectors. nr_sectors is more of getting the equal bandwidth even when we are operating at different request sizes. So instead of operating in iops mode, if you operate in regular time based mode, you should get better results. Why are you not using regular time based fairness mode? Thanks Vivek ^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: Re: Re: cfq-iosched.c:Use cfqq->nr_sectors in charge the vdisktime 2011-03-31 15:46 ` Lina Lu 2011-03-31 19:46 ` Vivek Goyal @ 2011-04-01 14:59 ` Lina Lu 2011-04-01 15:22 ` Vivek Goyal 1 sibling, 1 reply; 6+ messages in thread From: Lina Lu @ 2011-04-01 14:59 UTC (permalink / raw) To: Vivek Goyal; +Cc: linux kernel mailing list On 2011-04-01 03:47:18, Vivek Goyal wrote: > On Thu, Mar 31, 2011 at 11:46:37PM +0800, Lina Lu wrote: > > On 2011-03-30 23:54:34, Vivek Goyal wrote: > > [..] > > > > Here is 20 sec backtrace: > > http://www.fileden.com/files/2010/9/9/2965145/cfq_log.tar.gz > > > > This time, I set two IO pid with weight 100, and the device is in iops_mod. > > How did you put device in iops mode? What's the device you are using and > what kind of configuration dm-0 and dm-1 are in. I echo 0 to /sys/block/sdb/queue/iosched/slice_idle to put the device in iops mod. Here is the dmsetup table: sdbtest-2: 0 2097152 linear 8:16 23068672 sdbtest-1: 0 2097152 linear 8:16 20971520 Device dm-0 is sdbtest-1, and dm-1 is sdbtest-2. They are all linear logic devices on sdb. > > > linux-kzr4:/home/blkio # cat tst1/blkio.weight > > 100 > > linux-kzr4:/home/blkio # cat tst2/blkio.weight > > 100 > > > > iostat: > > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > > dm-0 0.00 0.00 855.50 0.00 3.34 0.00 8.00 0.82 1.06 0.95 81.70 > > dm-1 0.00 0.00 844.00 0.00 26.38 0.00 64.00 0.83 0.98 0.98 82.60 > > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > > dm-0 0.00 0.00 840.00 0.00 3.28 0.00 8.00 0.90 0.95 1.07 89.55 > > dm-1 0.00 0.00 794.00 0.00 24.81 0.00 64.00 0.87 1.10 1.10 87.00 > > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > > dm-0 0.00 0.00 596.50 0.00 2.33 0.00 8.00 0.96 1.77 1.61 95.80 > > dm-1 0.00 0.00 626.00 0.00 19.56 0.00 64.00 0.94 1.48 1.50 93.70 > > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > > dm-0 0.00 0.00 815.50 0.00 3.19 0.00 8.00 0.81 0.83 1.00 81.40 > > dm-1 0.00 0.00 828.50 0.00 25.89 0.00 64.00 0.77 0.95 0.93 77.45 > > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > > dm-0 0.00 0.00 910.50 0.00 3.56 0.00 8.00 0.82 1.00 0.90 82.15 > > dm-1 0.00 0.00 845.00 0.00 26.41 0.00 64.00 0.81 0.96 0.96 80.95 > > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > > dm-0 0.00 0.00 928.86 0.00 3.63 0.00 8.00 0.79 0.90 0.86 79.45 > > dm-1 0.00 0.00 848.26 0.00 26.51 0.00 64.00 0.65 0.77 0.77 65.17 > > > > >From the result, we can see that the iops match the weight value very well, but > > the rMB/s are not the same as they has different avgrq-sz. > > > > If I use the following patch, the rMB/s will be more accuracy. > > > > --- block/cfq-iosched.c 2011-03-31 23:43:55.000000000 +0800 > > +++ block/cfq-iosched.c 2011-03-31 23:44:30.000000000 +0800 > > @@ -951,7 +951,7 @@ > > used_sl = charge = cfq_cfqq_slice_usage(cfqq); > > > > if (iops_mode(cfqd)) > > - charge = cfqq->slice_dispatch; > > + charge = cfqq->nr_sectors; > > In IOPS mode we calculate the number of IOPS (that is number of requests > dispatched) and not number of sectors. nr_sectors is more of getting > the equal bandwidth even when we are operating at different request sizes. > So instead of operating in iops mode, if you operate in regular time > based mode, you should get better results. > > Why are you not using regular time based fairness mode? > I did the same test in regular time based fairness mode without the above patch. Here is iostat result: Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util dm-0 0.00 0.00 1813.00 0.00 7.08 0.00 8.00 0.81 0.42 0.45 81.40 dm-1 0.00 0.00 627.00 0.00 19.59 0.00 64.00 0.92 1.61 1.47 92.20 Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util dm-0 0.00 0.00 1799.00 0.00 7.03 0.00 8.00 0.80 0.44 0.44 80.00 dm-1 0.00 0.00 660.00 0.00 20.62 0.00 64.00 0.95 1.44 1.43 94.70 Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util dm-0 0.00 0.00 1875.00 0.00 7.32 0.00 8.00 0.68 0.39 0.36 67.60 dm-1 0.00 0.00 540.00 0.00 16.88 0.00 64.00 0.94 1.59 1.75 94.50 Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util dm-0 0.00 0.00 1494.06 0.00 5.84 0.00 8.00 0.73 0.45 0.49 73.27 dm-1 0.00 0.00 688.12 0.00 21.50 0.00 64.00 0.90 1.44 1.31 90.40 Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util dm-0 0.00 0.00 2079.00 0.00 8.12 0.00 8.00 0.80 0.41 0.38 79.50 dm-1 0.00 0.00 623.00 0.00 19.47 0.00 64.00 0.94 1.43 1.50 93.70 Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util dm-0 0.00 0.00 1991.00 0.00 7.78 0.00 8.00 0.87 0.44 0.44 86.80 dm-1 0.00 0.00 708.00 0.00 22.12 0.00 64.00 0.89 1.25 1.26 89.30 If I apply the above patch, and test in iops mode, the bandwidth will be equal. Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util dm-0 0.00 0.00 2579.00 0.00 10.07 0.00 8.00 0.92 0.35 0.36 91.80 dm-1 0.00 0.00 253.00 0.00 7.91 0.00 64.00 0.98 3.93 3.88 98.10 Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util dm-0 0.00 0.00 2394.00 0.00 9.35 0.00 8.00 0.93 0.40 0.39 93.00 dm-1 0.00 0.00 326.00 0.00 10.19 0.00 64.00 0.91 2.41 2.80 91.30 Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util dm-0 0.00 0.00 2339.00 0.00 9.14 0.00 8.00 0.91 0.37 0.39 90.50 dm-1 0.00 0.00 267.00 0.00 8.34 0.00 64.00 0.97 4.10 3.63 96.90 Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util dm-0 0.00 0.00 2298.00 0.00 8.98 0.00 8.00 0.59 0.25 0.26 59.00 dm-1 0.00 0.00 286.00 0.00 8.94 0.00 64.00 0.98 3.43 3.43 98.10 Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util dm-0 0.00 0.00 2298.00 0.00 8.98 0.00 8.00 0.37 0.18 0.16 37.00 dm-1 0.00 0.00 292.00 0.00 9.12 0.00 64.00 0.98 2.83 3.35 97.80 But it seens the total performance is lower. Thanks Lina f ^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: Re: Re: cfq-iosched.c:Use cfqq->nr_sectors in charge the vdisktime 2011-04-01 14:59 ` Lina Lu @ 2011-04-01 15:22 ` Vivek Goyal 0 siblings, 0 replies; 6+ messages in thread From: Vivek Goyal @ 2011-04-01 15:22 UTC (permalink / raw) To: Lina Lu; +Cc: linux kernel mailing list On Fri, Apr 01, 2011 at 10:59:52PM +0800, Lina Lu wrote: > On 2011-04-01 03:47:18, Vivek Goyal wrote: > > On Thu, Mar 31, 2011 at 11:46:37PM +0800, Lina Lu wrote: > > > On 2011-03-30 23:54:34, Vivek Goyal wrote: > > > [..] > > > > > > Here is 20 sec backtrace: > > > http://www.fileden.com/files/2010/9/9/2965145/cfq_log.tar.gz > > > > > > This time, I set two IO pid with weight 100, and the device is in iops_mod. > > > > How did you put device in iops mode? What's the device you are using and > > what kind of configuration dm-0 and dm-1 are in. > > I echo 0 to /sys/block/sdb/queue/iosched/slice_idle to put the device in iops mod. > > Here is the dmsetup table: > sdbtest-2: 0 2097152 linear 8:16 23068672 > sdbtest-1: 0 2097152 linear 8:16 20971520 > > Device dm-0 is sdbtest-1, and dm-1 is sdbtest-2. They are all linear logic devices > on sdb. > > > > > > linux-kzr4:/home/blkio # cat tst1/blkio.weight > > > 100 > > > linux-kzr4:/home/blkio # cat tst2/blkio.weight > > > 100 > > > > > > iostat: > > > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > > > dm-0 0.00 0.00 855.50 0.00 3.34 0.00 8.00 0.82 1.06 0.95 81.70 > > > dm-1 0.00 0.00 844.00 0.00 26.38 0.00 64.00 0.83 0.98 0.98 82.60 > > > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > > > dm-0 0.00 0.00 840.00 0.00 3.28 0.00 8.00 0.90 0.95 1.07 89.55 > > > dm-1 0.00 0.00 794.00 0.00 24.81 0.00 64.00 0.87 1.10 1.10 87.00 > > > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > > > dm-0 0.00 0.00 596.50 0.00 2.33 0.00 8.00 0.96 1.77 1.61 95.80 > > > dm-1 0.00 0.00 626.00 0.00 19.56 0.00 64.00 0.94 1.48 1.50 93.70 > > > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > > > dm-0 0.00 0.00 815.50 0.00 3.19 0.00 8.00 0.81 0.83 1.00 81.40 > > > dm-1 0.00 0.00 828.50 0.00 25.89 0.00 64.00 0.77 0.95 0.93 77.45 > > > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > > > dm-0 0.00 0.00 910.50 0.00 3.56 0.00 8.00 0.82 1.00 0.90 82.15 > > > dm-1 0.00 0.00 845.00 0.00 26.41 0.00 64.00 0.81 0.96 0.96 80.95 > > > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > > > dm-0 0.00 0.00 928.86 0.00 3.63 0.00 8.00 0.79 0.90 0.86 79.45 > > > dm-1 0.00 0.00 848.26 0.00 26.51 0.00 64.00 0.65 0.77 0.77 65.17 > > > > > > >From the result, we can see that the iops match the weight value very well, but > > > the rMB/s are not the same as they has different avgrq-sz. > > > > > > If I use the following patch, the rMB/s will be more accuracy. > > > > > > --- block/cfq-iosched.c 2011-03-31 23:43:55.000000000 +0800 > > > +++ block/cfq-iosched.c 2011-03-31 23:44:30.000000000 +0800 > > > @@ -951,7 +951,7 @@ > > > used_sl = charge = cfq_cfqq_slice_usage(cfqq); > > > > > > if (iops_mode(cfqd)) > > > - charge = cfqq->slice_dispatch; > > > + charge = cfqq->nr_sectors; > > > > In IOPS mode we calculate the number of IOPS (that is number of requests > > dispatched) and not number of sectors. nr_sectors is more of getting > > the equal bandwidth even when we are operating at different request sizes. > > So instead of operating in iops mode, if you operate in regular time > > based mode, you should get better results. > > > > Why are you not using regular time based fairness mode? > > > > I did the same test in regular time based fairness mode without the above patch. > > Here is iostat result: > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > dm-0 0.00 0.00 1813.00 0.00 7.08 0.00 8.00 0.81 0.42 0.45 81.40 > dm-1 0.00 0.00 627.00 0.00 19.59 0.00 64.00 0.92 1.61 1.47 92.20 > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > dm-0 0.00 0.00 1799.00 0.00 7.03 0.00 8.00 0.80 0.44 0.44 80.00 > dm-1 0.00 0.00 660.00 0.00 20.62 0.00 64.00 0.95 1.44 1.43 94.70 > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > dm-0 0.00 0.00 1875.00 0.00 7.32 0.00 8.00 0.68 0.39 0.36 67.60 > dm-1 0.00 0.00 540.00 0.00 16.88 0.00 64.00 0.94 1.59 1.75 94.50 > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > dm-0 0.00 0.00 1494.06 0.00 5.84 0.00 8.00 0.73 0.45 0.49 73.27 > dm-1 0.00 0.00 688.12 0.00 21.50 0.00 64.00 0.90 1.44 1.31 90.40 > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > dm-0 0.00 0.00 2079.00 0.00 8.12 0.00 8.00 0.80 0.41 0.38 79.50 > dm-1 0.00 0.00 623.00 0.00 19.47 0.00 64.00 0.94 1.43 1.50 93.70 > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > dm-0 0.00 0.00 1991.00 0.00 7.78 0.00 8.00 0.87 0.44 0.44 86.80 > dm-1 0.00 0.00 708.00 0.00 22.12 0.00 64.00 0.89 1.25 1.26 89.30 > > If I apply the above patch, and test in iops mode, the bandwidth will be equal. > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > dm-0 0.00 0.00 2579.00 0.00 10.07 0.00 8.00 0.92 0.35 0.36 91.80 > dm-1 0.00 0.00 253.00 0.00 7.91 0.00 64.00 0.98 3.93 3.88 98.10 > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > dm-0 0.00 0.00 2394.00 0.00 9.35 0.00 8.00 0.93 0.40 0.39 93.00 > dm-1 0.00 0.00 326.00 0.00 10.19 0.00 64.00 0.91 2.41 2.80 91.30 > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > dm-0 0.00 0.00 2339.00 0.00 9.14 0.00 8.00 0.91 0.37 0.39 90.50 > dm-1 0.00 0.00 267.00 0.00 8.34 0.00 64.00 0.97 4.10 3.63 96.90 > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > dm-0 0.00 0.00 2298.00 0.00 8.98 0.00 8.00 0.59 0.25 0.26 59.00 > dm-1 0.00 0.00 286.00 0.00 8.94 0.00 64.00 0.98 3.43 3.43 98.10 > Device: rrqm/s wrqm/s r/s w/s rMB/s wMB/s avgrq-sz avgqu-sz await svctm %util > dm-0 0.00 0.00 2298.00 0.00 8.98 0.00 8.00 0.37 0.18 0.16 37.00 > dm-1 0.00 0.00 292.00 0.00 9.12 0.00 64.00 0.98 2.83 3.35 97.80 > > But it seens the total performance is lower. Lina, We really don't have any equal bandwidth mode. In time mode, every queue is given specific time slice of disk. If a group is doing bigger size IO and can get higher bandwidth from disk in allotted time slice then it makes sense. That group made better use of its time slice. Trying to penalize the group which is doing bigger size IO because some other group is doing small size IO does not make much sense to me. Similar thing is true for sequential vs seeky load. If you run sequential process on dm-0 and seeky process on dm-1, you will see overall bandwdith difference. If you want to give higher bandwidth to the group doing small size IO, just bump up its weight and you should get similar results as you are getting with your patch applied. Thanks Vivek ^ permalink raw reply [flat|nested] 6+ messages in thread
end of thread, other threads:[~2011-04-01 15:22 UTC | newest] Thread overview: 6+ messages (download: mbox.gz follow: Atom feed -- links below jump to the message on this page -- 2011-03-30 15:23 cfq-iosched.c:Use cfqq->nr_sectors in charge the vdisktime Lina Lu 2011-03-30 15:54 ` Vivek Goyal 2011-03-31 15:46 ` Lina Lu 2011-03-31 19:46 ` Vivek Goyal 2011-04-01 14:59 ` Lina Lu 2011-04-01 15:22 ` Vivek Goyal
This is a public inbox, see mirroring instructions for how to clone and mirror all data and code used for this inbox