Date: Fri, 31 Jul 2009 09:13:59 -0400
From: Vivek Goyal <vgoyal@redhat.com>
To: Gui Jianfeng <guijianfeng@cn.fujitsu.com>
Cc: linux-kernel@vger.kernel.org, containers@lists.linux-foundation.org,
       dm-devel@redhat.com, jens.axboe@oracle.com, nauman@google.com,
       dpshah@google.com, ryov@valinux.co.jp, balbir@linux.vnet.ibm.com,
       righi.andrea@gmail.com, lizf@cn.fujitsu.com, mikew@google.com,
       fchecconi@gmail.com, paolo.valente@unimore.it, fernando@oss.ntt.co.jp,
       s-uchida@ap.jp.nec.com, taka@valinux.co.jp, jmoyer@redhat.com,
       dhaval@linux.vnet.ibm.com, m-ikeda@ds.jp.nec.com, agk@redhat.com,
       akpm@linux-foundation.org, peterz@infradead.org
Subject: Re: [RFC] IO scheduler based IO controller V7
Message-ID: <20090731131359.GA3668@redhat.com>
References: <1248467274-32073-1-git-send-email-vgoyal@redhat.com> <4A727F6F.9010005@cn.fujitsu.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <4A727F6F.9010005@cn.fujitsu.com>
User-Agent: Mutt/1.5.18 (2008-05-17)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 18980
Lines: 440

On Fri, Jul 31, 2009 at 01:21:51PM +0800, Gui Jianfeng wrote:
> Hi Vivek,
> 
> Here are some test results for normal reads and write for IO Controller V7 by fio.
> Tested with "fairness == 0". It seems performance gets better comparing with V6.
> 
> Mode         Normal read   |   Random read   |   Normal write   |   Random write  |  Direct read  |  Direct Write
> 
> 2.6.31-rc1   71,613KiB/s       3,606KiB/s        66,250KiB/s        9,420KiB/s       51,535KiB/s     55,752KiB/s
> 
> V7           70,540KiB/s       3,551KiB/s        64,548KiB/s        9,677KiB/s       53,530KiB/s     54,145KiB/s
> 
> Performance  -1.5%             -1.5%             -2.6%              +2.7%            +3.9%           -2.9%
> 

Thanks Gui. Can you also try V7 with CONFIG_TRACK_ASYNC_CONTEXT=n. I tried
that and I got better results for buffered writes.

In my testing I still see some performance regression for buffered writes
which goes away if I disable group io scheduling and just use flat mode.

I will spend more time to find out where it is coming from.

Thanks
Vivek


> 
> Vivek Goyal wrote:
> > Hi All,
> > 
> > Here is the V7 of the IO controller patches generated on top of 2.6.31-rc4.
> > 
> > For ease of patching, a consolidated patch is available here.
> > 
> > http://people.redhat.com/~vgoyal/io-controller/io-scheduler-based-io-controller-v7.patch
> > 
> > Previous versions of the patches was posted here.
> > 
> > (V1) http://lkml.org/lkml/2009/3/11/486
> > (V2) http://lkml.org/lkml/2009/5/5/275
> > (V3) http://lkml.org/lkml/2009/5/26/472
> > (V4) http://lkml.org/lkml/2009/6/8/580
> > (V5) http://lkml.org/lkml/2009/6/19/279
> > (V6) http://lkml.org/lkml/2009/7/2/369
> > 
> > Changes from V6
> > ===============
> > - Introduced the notion of group_idling where we idle for next request to
> >   come from the same group before we expire it. It is along the lines of
> >   cfq's slice_idle thing to provide fairness. Switching to group idling
> >   now helps in the sense that we don't have to rely whether queue idling
> >   was turned on or not by CFQ. It becomes too much of debugging pain with
> >   different work loads and different kind of storage media. Introduction
> >   of group_idle should help.
> > 
> > - Moved some of the code like dynamic queue idling update, arming queue
> >   idling timer, keeping track of average think time etc back to CFQ. With
> >   group idling we don't need it now. Reduce the amount of change.
> > 
> > - Enabled cfq's close cooperator functionality in groups. So far this worked
> >   only in root group. Now it should work in non-root groups also.
> > 
> > - Got rid of the patch where we calculated disk time based on average disk
> >   rate in some circumstances. It was giving bad numbers in early queue
> >   deletion cases. Also did not think that it was helping a lot. Remvoed it
> >   for the time being.
> >  
> > - Added an experimental patch to map sync requests using bio tracking info and
> >   not task context. This is only for noop, deadline and AS.
> > 
> > - Got rid of experimental patch of idling for async queues. Don't think it
> >   was helping.
> > 
> > - Got rid of wait_busy and wait_busy_done logic from queue. Instead
> >   implemented it for groups.
> > 
> > - Introduced oom_ioq to accomodate oom_cfqq change recently.
> > 
> > - Broke-up elv_init_ioq() fn into smaller functions. It had 7 arguments and
> >   looked complicated.
> > 
> > - Fixed a bug in blk_queue_io_group_congested(). Thanks to Munehiro Ikeda.
> > 
> > - Merged gui's patch to fix the cgroup file format issue.
> > 
> > - Merged gui's patch to update per group congestion limit when
> >   q->nr_group_requests is updated.
> > 
> > - Fixed a bug where close cooperation will not work if we wait for all the
> >   requests to finish from previous queue.
> > 
> > - Fixed group deletion accouting where deletion from idle tree were also
> >   appearing in the log.
> > 
> > - Got rid of busy_rt_queues infrastructure.
> > 
> > - Got rid of elv_ioq_request_dispatched(). An helper function just to
> >   increment a variable.
> >   
> > Limitations
> > ===========
> > 
> > - This IO controller provides the bandwidth control at the IO scheduler
> >   level (leaf node in stacked hiearchy of logical devices). So there can
> >   be cases (depending on configuration) where application does not see
> >   proportional BW division at higher logical level device.
> > 
> >   LWN has written an article about the issue here.
> > 
> > 	http://lwn.net/Articles/332839/
> > 
> > How to solve the issue of fairness at higher level logical devices
> > ==================================================================
> > (Do we really need it? That's not where the contention for resources is.)
> > 
> > Couple of suggestions have come forward.
> > 
> > - Implement IO control at IO scheduler layer and then with the help of
> >   some daemon, adjust the weight on underlying devices dynamiclly, depending
> >   on what kind of BW gurantees are to be achieved at higher level logical
> >   block devices.
> > 
> > - Also implement a higher level IO controller along with IO scheduler
> >   based controller and let user choose one depending on his needs.
> > 
> >   A higher level controller does not know about the assumptions/policies
> >   of unerldying IO scheduler, hence it has the potential to break down
> >   the IO scheduler's policy with-in cgroup. A lower level controller
> >   can work with IO scheduler much more closely and efficiently.
> >  
> > Other active IO controller developments
> > =======================================
> > 
> > IO throttling
> > -------------
> > 
> >   This is a max bandwidth controller and not the proportional one. Secondly
> >   it is a second level controller which can break the IO scheduler's
> >   policy/assumtions with-in cgroup. 
> > 
> > dm-ioband
> > ---------
> > 
> >  This is a proportional bandwidth controller implemented as device mapper
> >  driver. It is also a second level controller which can break the
> >  IO scheduler's policy/assumptions with-in cgroup.
> > 
> > TODO
> > ====
> > - code cleanups, testing, bug fixing, optimizations, benchmarking etc...
> > 
> > Testing
> > =======
> > 
> > I have been able to do some testing as follows. All my testing is with ext3
> > file system with a SATA drive which supports queue depth of 31.
> > 
> > Test1 (Isolation between two KVM virtual machines)
> > ==================================================
> > Created two KVM virtual machines. Partitioned a disk on host in two partitions
> > and gave one partition to each virtual machine. Put both the virtual machines
> > in two different cgroup of weight 1000 and 500 each. Virtual machines created
> > ext3 file system on the partitions exported from host and did buffered writes.
> > Host seems writes as synchronous and virtual machine with higher weight gets
> > double the disk time of virtual machine of lower weight. Used deadline
> > scheduler in this test case.
> > 
> > Some more details about configuration are in documentation patch.
> > 
> > Test2 (Fairness for synchronous reads)
> > ======================================
> > - Two dd in two cgroups with cgrop weights 1000 and 500. Ran two "dd" in those
> >   cgroups (With CFQ scheduler and /sys/block/<device>/queue/fairness = 1)
> > 
> >   Higher weight dd finishes first and at that point of time my script takes
> >   care of reading cgroup files io.disk_time and io.disk_sectors for both the
> >   groups and display the results.
> > 
> >   dd if=/mnt/$BLOCKDEV/zerofile1 of=/dev/null &
> >   dd if=/mnt/$BLOCKDEV/zerofile2 of=/dev/null &
> > 
> >   234179072 bytes (234 MB) copied, 3.9065 s, 59.9 MB/s
> >   234179072 bytes (234 MB) copied, 5.19232 s, 45.1 MB/s
> > 
> >   group1 time=8 16 2471 group1 sectors=8 16 457840
> >   group2 time=8 16 1220 group2 sectors=8 16 225736
> > 
> > First two fields in time and sectors statistics represent major and minor
> > number of the device. Third field represents disk time in milliseconds and
> > number of sectors transferred respectively.
> > 
> > This patchset tries to provide fairness in terms of disk time received. group1
> > got almost double of group2 disk time (At the time of first dd finish). These
> > time and sectors statistics can be read using io.disk_time and io.disk_sector
> > files in cgroup. More about it in documentation file.
> > 
> > Test3 (Reader Vs Buffered Writes)
> > ================================
> > Buffered writes can be problematic and can overwhelm readers, especially with
> > noop and deadline. IO controller can provide isolation between readers and
> > buffered (async) writers.
> > 
> > First I ran the test without io controller to see the severity of the issue.
> > Ran a hostile writer and then after 10 seconds started a reader and then
> > monitored the completion time of reader. Reader reads a 256 MB file. Tested
> > this with noop scheduler.
> > 
> > sample script
> > ------------
> > sync
> > echo 3 > /proc/sys/vm/drop_caches
> > time dd if=/dev/zero of=/mnt/sdb/reader-writer-zerofile bs=4K count=2097152
> > conv=fdatasync &
> > sleep 10
> > time dd if=/mnt/sdb/256M-file of=/dev/null &
> > 
> > Results
> > -------
> > 8589934592 bytes (8.6 GB) copied, 106.045 s, 81.0 MB/s (Writer)
> > 268435456 bytes (268 MB) copied, 96.5237 s, 2.8 MB/s (Reader)
> > 
> > Now it was time to test io controller whether it can provide isolation between
> > readers and writers with noop. I created two cgroups of weight 1000 each and
> > put reader in group1 and writer in group 2 and ran the test again. Upon
> > comletion of reader, my scripts read io.dis_time and io.disk_group cgroup
> > files to get an estimate how much disk time each group got and how many
> > sectors each group did IO for. 
> > 
> > For more accurate accounting of disk time for buffered writes with queuing
> > hardware I had to set /sys/block/<disk>/queue/iosched/fairness to "1".
> > 
> > sample script
> > -------------
> > echo $$ > /cgroup/bfqio/test2/tasks
> > dd if=/dev/zero of=/mnt/$BLOCKDEV/testzerofile bs=4K count=2097152 &
> > sleep 10
> > echo noop > /sys/block/$BLOCKDEV/queue/scheduler
> > echo  1 > /sys/block/$BLOCKDEV/queue/iosched/fairness
> > echo $$ > /cgroup/bfqio/test1/tasks
> > dd if=/mnt/$BLOCKDEV/256M-file of=/dev/null &
> > wait $!
> > # Some code for reading cgroup files upon completion of reader.
> > -------------------------
> > 
> > Results
> > =======
> > 268435456 bytes (268 MB) copied, 6.65819 s, 40.3 MB/s (Reader) 
> > 
> > group1 time=8 16 3063	group1 sectors=8 16 524808
> > group2 time=8 16 3071	group2 sectors=8 16 441752
> > 
> > Note, reader finishes now much lesser time and both group1 and group2
> > got almost 3 seconds of disk time. Hence io-controller provides isolation
> > from buffered writes.
> > 
> > Test4 (AIO)
> > ===========
> > 
> > AIO reads
> > -----------
> > Set up two fio, AIO read jobs in two cgroup with weight 1000 and 500
> > respectively. I am using cfq scheduler. Following are some lines from my test
> > script.
> > 
> > ---------------------------------------------------------------
> > echo 1000 > /cgroup/bfqio/test1/io.weight
> > echo 500 > /cgroup/bfqio/test2/io.weight
> > 
> > fio_args="--ioengine=libaio --rw=read --size=512M --direct=1"
> > echo 1 > /sys/block/$BLOCKDEV/queue/iosched/fairness
> > 
> > echo $$ > /cgroup/bfqio/test1/tasks
> > fio $fio_args --name=test1 --directory=/mnt/$BLOCKDEV/fio1/
> > --output=/mnt/$BLOCKDEV/fio1/test1.log
> > --exec_postrun="../read-and-display-group-stats.sh $maj_dev $minor_dev" &
> > 
> > echo $$ > /cgroup/bfqio/test2/tasks
> > fio $fio_args --name=test2 --directory=/mnt/$BLOCKDEV/fio2/
> > --output=/mnt/$BLOCKDEV/fio2/test2.log &
> > ----------------------------------------------------------------
> > 
> > test1 and test2 are two groups with weight 1000 and 500 respectively.
> > "read-and-display-group-stats.sh" is one small script which reads the
> > test1 and test2 cgroup files to determine how much disk time each group
> > got till first fio job finished.
> > 
> > Results
> > ------
> > test1 statistics: time=8 16 22403   sectors=8 16 1049640
> > test2 statistics: time=8 16 11400   sectors=8 16 552864
> > 
> > Above shows that by the time first fio (higher weight), finished, group
> > test1 got 22403 ms of disk time and group test2 got 11400 ms of disk time.
> > similarly the statistics for number of sectors transferred are also shown.
> > 
> > Note that disk time given to group test1 is almost double of group2 disk
> > time.
> > 
> > AIO writes
> > ----------
> > Set up two fio, AIO direct write jobs in two cgroup with weight 1000 and 500
> > respectively. I am using cfq scheduler. Following are some lines from my test
> > script.
> > 
> > ------------------------------------------------
> > echo 1000 > /cgroup/bfqio/test1/io.weight
> > echo 500 > /cgroup/bfqio/test2/io.weight
> > fio_args="--ioengine=libaio --rw=write --size=512M --direct=1"
> > 
> > echo 1 > /sys/block/$BLOCKDEV/queue/iosched/fairness
> > 
> > echo $$ > /cgroup/bfqio/test1/tasks
> > fio $fio_args --name=test1 --directory=/mnt/$BLOCKDEV/fio1/
> > --output=/mnt/$BLOCKDEV/fio1/test1.log
> > --exec_postrun="../read-and-display-group-stats.sh $maj_dev $minor_dev" &
> > 
> > echo $$ > /cgroup/bfqio/test2/tasks
> > fio $fio_args --name=test2 --directory=/mnt/$BLOCKDEV/fio2/
> > --output=/mnt/$BLOCKDEV/fio2/test2.log &
> > -------------------------------------------------
> > 
> > test1 and test2 are two groups with weight 1000 and 500 respectively.
> > "read-and-display-group-stats.sh" is one small script which reads the
> > test1 and test2 cgroup files to determine how much disk time each group
> > got till first fio job finished.
> > 
> > Following are the results.
> > 
> > test1 statistics: time=8 16 29085   sectors=8 16 1049656
> > test2 statistics: time=8 16 14652   sectors=8 16 516728
> > 
> > Above shows that by the time first fio (higher weight), finished, group
> > test1 got 28085 ms of disk time and group test2 got 14652 ms of disk time.
> > similarly the statistics for number of sectors transferred are also shown.
> > 
> > Note that disk time given to group test1 is almost double of group2 disk
> > time.
> > 
> > Test5 (Fairness for async writes, Buffered Write Vs Buffered Write)
> > ===================================================================
> > Fairness for async writes is tricky and biggest reason is that async writes
> > are cached in higher layers (page cahe) as well as possibly in file system
> > layer also (btrfs, xfs etc), and are dispatched to lower layers not necessarily
> > in proportional manner.
> > 
> > For example, consider two dd threads reading /dev/zero as input file and doing
> > writes of huge files. Very soon we will cross vm_dirty_ratio and dd thread will
> > be forced to write out some pages to disk before more pages can be dirtied. But
> > not necessarily dirty pages of same thread are picked. It can very well pick
> > the inode of lesser priority dd thread and do some writeout. So effectively
> > higher weight dd is doing writeouts of lower weight dd pages and we don't see
> > service differentation.
> > 
> > IOW, the core problem with async write fairness is that higher weight thread
> > does not throw enought IO traffic at IO controller to keep the queue
> > continuously backlogged. In my testing, there are many .2 to .8 second
> > intervals where higher weight queue is empty and in that duration lower weight
> > queue get lots of job done giving the impression that there was no service
> > differentiation.
> > 
> > In summary, from IO controller point of view async writes support is there.
> > Because page cache has not been designed in such a manner that higher 
> > prio/weight writer can do more write out as compared to lower prio/weight
> > writer, gettting service differentiation is hard and it is visible in some
> > cases and not visible in some cases.
> > 
> > Do we really care that much for fairness among two writer cgroups? One can
> > choose to do direct writes or sync writes if fairness for writes really
> > matters for him.
> > 
> > Following is the only case where it is hard to ensure fairness between cgroups.
> > 
> > - Buffered writes Vs Buffered Writes.
> > 
> > So to test async writes I created two partitions on a disk and created ext3
> > file systems on both the partitions.  Also created two cgroups and generated
> > lots of write traffic in two cgroups (50 fio threads) and watched the disk
> > time statistics in respective cgroups at the interval of 2 seconds. Thanks to
> > ryo tsuruta for the test case.
> > 
> > *****************************************************************
> > sync
> > echo 3 > /proc/sys/vm/drop_caches
> > 
> > fio_args="--size=64m --rw=write --numjobs=50 --group_reporting"
> > 
> > echo $$ > /cgroup/bfqio/test1/tasks
> > fio $fio_args --name=test1 --directory=/mnt/sdd1/fio/ --output=/mnt/sdd1/fio/test1.log &
> > 
> > echo $$ > /cgroup/bfqio/test2/tasks
> > fio $fio_args --name=test2 --directory=/mnt/sdd2/fio/ --output=/mnt/sdd2/fio/test2.log &
> > *********************************************************************** 
> > 
> > And watched the disk time and sector statistics for the both the cgroups
> > every 2 seconds using a script. How is snippet from output.
> > 
> > test1 statistics: time=8 48 1315   sectors=8 48 55776 dq=8 48 1
> > test2 statistics: time=8 48 633   sectors=8 48 14720 dq=8 48 2
> > 
> > test1 statistics: time=8 48 5586   sectors=8 48 339064 dq=8 48 2
> > test2 statistics: time=8 48 2985   sectors=8 48 146656 dq=8 48 3
> > 
> > test1 statistics: time=8 48 9935   sectors=8 48 628728 dq=8 48 3
> > test2 statistics: time=8 48 5265   sectors=8 48 278688 dq=8 48 4
> > 
> > test1 statistics: time=8 48 14156   sectors=8 48 932488 dq=8 48 6
> > test2 statistics: time=8 48 7646   sectors=8 48 412704 dq=8 48 7
> > 
> > test1 statistics: time=8 48 18141   sectors=8 48 1231488 dq=8 48 10
> > test2 statistics: time=8 48 9820   sectors=8 48 548400 dq=8 48 8
> > 
> > test1 statistics: time=8 48 21953   sectors=8 48 1485632 dq=8 48 13
> > test2 statistics: time=8 48 12394   sectors=8 48 698288 dq=8 48 10
> > 
> > test1 statistics: time=8 48 25167   sectors=8 48 1705264 dq=8 48 13
> > test2 statistics: time=8 48 14042   sectors=8 48 817808 dq=8 48 10
> > 
> > First two fields in time and sectors statistics represent major and minor
> > number of the device. Third field represents disk time in milliseconds and
> > number of sectors transferred respectively.
> > 
> > So disk time consumed by group1 is almost double of group2 in this case.
> > 
> > Your feedback is welcome.
> > 
> > Thanks
> > Vivek
> > 
> > 
> > 
> 
> -- 
> Regards
> Gui Jianfeng
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/