Re: IO scheduler based IO controller V10

On Thu, Sep 24, 2009 at 02:33:15PM -0700, Andrew Morton wrote:
On Thu, 24 Sep 2009 15:25:04 -0400
Vivek Goyal <vgoyal@xxxxxxxxxx> wrote:


Hi All,

Here is the V10 of the IO controller patches generated on top of 2.6.31.


Thanks for the writeup. It really helps and is most worthwhile for a
project of this importance, size and complexity.



What problem are we trying to solve
===================================
Provide group IO scheduling feature in Linux along the lines of other resource
controllers like cpu.

IOW, provide facility so that a user can group applications using cgroups and
control the amount of disk time/bandwidth received by a group based on its
weight.

How to solve the problem
=========================

Different people have solved the issue differetnly. So far looks it looks
like we seem to have following two core requirements when it comes to
fairness at group level.

- Control bandwidth seen by groups.
- Control on latencies when a request gets backlogged in group.

At least there are now three patchsets available (including this one).

IO throttling
-------------
This is a bandwidth controller which keeps track of IO rate of a group and
throttles the process in the group if it exceeds the user specified limit.

dm-ioband
---------
This is a proportional bandwidth controller implemented as device mapper
driver and provides fair access in terms of amount of IO done (not in terms
of disk time as CFQ does).

So one will setup one or more dm-ioband devices on top of physical/logical
block device, configure the ioband device and pass information like grouping
etc. Now this device will keep track of bios flowing through it and control
the flow of bios based on group policies.

IO scheduler based IO controller
--------------------------------
Here we have viewed the problem of IO contoller as hierarchical group
scheduling (along the lines of CFS group scheduling) issue. Currently one can
view linux IO schedulers as flat where there is one root group and all the IO
belongs to that group.

This patchset basically modifies IO schedulers to also support hierarchical
group scheduling. CFQ already provides fairness among different processes. I
have extended it support group IO schduling. Also took some of the code out
of CFQ and put in a common layer so that same group scheduling code can be
used by noop, deadline and AS to support group scheduling.

Pros/Cons
=========
There are pros and cons to each of the approach. Following are some of the
thoughts.

Max bandwidth vs proportional bandwidth
---------------------------------------
IO throttling is a max bandwidth controller and not a proportional one.
Additionaly it provides fairness in terms of amount of IO done (and not in
terms of disk time as CFQ does).

Personally, I think that proportional weight controller is useful to more
people than just max bandwidth controller. In addition, IO scheduler based
controller can also be enhanced to do max bandwidth control. So it can
satisfy wider set of requirements.

Fairness in terms of disk time vs size of IO
---------------------------------------------
An higher level controller will most likely be limited to providing fairness
in terms of size/number of IO done and will find it hard to provide fairness
in terms of disk time used (as CFQ provides between various prio levels). This
is because only IO scheduler knows how much disk time a queue has used and
information about queues and disk time used is not exported to higher
layers.

So a seeky application will still run away with lot of disk time and bring
down the overall throughput of the the disk.

But that's only true if the thing is poorly implemented.

A high-level controller will need some view of the busyness of the
underlying device(s). That could be "proportion of idle time", or
"average length of queue" or "average request latency" or some mix of
these or something else altogether.

But these things are simple to calculate, and are simple to feed back
to the higher-level controller and probably don't require any changes
to to IO scheduler at all, which is a great advantage.


And I must say that high-level throttling based upon feedback from
lower layers seems like a much better model to me than hacking away in
the IO scheduler layer. Both from an implementation point of view and
from a "we can get it to work on things other than block devices" point
of view.


Hi Andrew,

Few thoughts.

- A higher level throttling approach suffers from the issue of unfair
throttling. So if there are multiple tasks in the group, who do we
throttle and how do we make sure that we did throttling in proportion
to the prio of tasks. Andrea's IO throttling implementation suffered
from these issues. I had run some tests where RT and BW tasks were
getting same BW with-in group or tasks of different prio were gettting
same BW.

Even if we figure a way out to do fair throttling with-in group, underlying
IO scheduler might not be CFQ at all and we should not have done so.

https://lists.linux-foundation.org/pipermail/containers/2009-May/017588.html

- Higher level throttling does not know where actually IO is going in
physical layer. So we might unnecessarily be throttling IO which are
going to same logical device but at the end of day to different physical
devices.

Agreed that some people will want that behavior, especially in the case
of max bandwidth control where one does not want to give you the BW
because you did not pay for it.

So higher level controller is good for max bw control but if it comes
to optimal usage of resources and do control only if needed, then it
probably is not the best thing.

About the feedback thing, I am not very sure. Are you saying that we will
run timed groups in higher layer and take feedback from underlying IO
scheduler about how much time a group consumed or something like that and
not do accounting in terms of size of IO?


Currently dm-ioband provides fairness in terms of number/size of IO.

Latencies and isolation between groups
--------------------------------------
An higher level controller is generally implementing a bandwidth throttling
solution where if a group exceeds either the max bandwidth or the proportional
share then throttle that group.

This kind of approach will probably not help in controlling latencies as it
will depend on underlying IO scheduler. Consider following scenario.

Assume there are two groups. One group is running multiple sequential readers
and other group has a random reader. sequential readers will get a nice 100ms
slice

Do you refer to each reader within group1, or to all readers? It would be
daft if each reader in group1 were to get 100ms.


All readers in the group should get 100ms each, both in IO throttling and
dm-ioband solution.

Higher level solutions are not keeping track of time slices. Time slices will
be allocated by CFQ which does not have any idea about grouping. Higher
level controller just keeps track of size of IO done at group level and
then run either a leaky bucket or token bucket algorithm.

IO throttling is a max BW controller, so it will not even care about what is
happening in other group. It will just be concerned with rate of IO in one
particular group and if we exceed specified limit, throttle it. So until and
unless sequential reader group hits it max bw limit, it will keep sending
reads down to CFQ, and CFQ will happily assign 100ms slices to readers.

dm-ioband will not try to choke the high throughput sequential reader group
for the slow random reader group because that would just kill the throughput
of rotational media. Every sequential reader will run for few ms and then
be throttled and this goes on. Disk will soon be seek bound.

each and then a random reader from group2 will get to dispatch the
request. So latency of this random reader will depend on how many sequential
readers are running in other group and that is a weak isolation between groups.

And yet that is what you appear to mean.

But surely nobody would do that - the 100ms would be assigned to and
distributed amongst all readers in group1?

Dividing 100ms to all the sequential readers might not be very good on
rotational media as each reader runs for small time and then seek happens.
This will increase number of seeks in the system. Think of 32 sequential
readers in the group and then each getting less than 3ms to run.

A better way probably is to give each queue 100ms in one run of group and
then switch group. Someting like following.

SR1 RR SR2 RR SR3 RR SR4 RR...

Now each sequential reader gets 100ms and disk is not seek bound at the
same time random reader latency limited by number of competing groups
and not by number of processes in the group. This is what IO scheduler
based IO controller is effectively doing currently.


When we control things at IO scheduler level, we assign one time slice to one
group and then pick next entity to run. So effectively after one time slice
(max 180ms, if prio 0 sequential reader is running), random reader in other
group will get to run. Hence we achieve better isolation between groups as
response time of process in a differnt group is generally not dependent on
number of processes running in competing group.

I don't understand why you're comparing this implementation with such
an obviously dumb competing design!

So a higher level solution is most likely limited to only shaping bandwidth
without any control on latencies.

Stacking group scheduler on top of CFQ can lead to issues
---------------------------------------------------------
IO throttling and dm-ioband both are second level controller. That is these
controllers are implemented in higher layers than io schedulers. So they
control the IO at higher layer based on group policies and later IO
schedulers take care of dispatching these bios to disk.

Implementing a second level controller has the advantage of being able to
provide bandwidth control even on logical block devices in the IO stack
which don't have any IO schedulers attached to these. But they can also
interefere with IO scheduling policy of underlying IO scheduler and change
the effective behavior. Following are some of the issues which I think
should be visible in second level controller in one form or other.

Prio with-in group
------------------
A second level controller can potentially interefere with behavior of
different prio processes with-in a group. bios are buffered at higher layer
in single queue and release of bios is FIFO and not proportionate to the
ioprio of the process. This can result in a particular prio level not
getting fair share.

That's an administrator error, isn't it? Should have put the
different-priority processes into different groups.


I am thinking in practice it probably will be a mix of priority in each
group. For example, consider a hypothetical scenario where two students
on a university server are given two cgroups of certain weights so that IO
done by these students are limited in case of contention. Now these students
might want to throw in a mix of priority workload in their respective cgroup.
Admin would not have any idea what priority process students are running in
respective cgroup.

Buffering at higher layer can delay read requests for more than slice idle
period of CFQ (default 8 ms). That means, it is possible that we are waiting
for a request from the queue but it is buffered at higher layer and then idle
timer will fire. It means that queue will losse its share at the same time
overall throughput will be impacted as we lost those 8 ms.

That sounds like a bug.


Actually this probably is a limitation of higher level controller. It most
likely is sitting so high in IO stack that it has no idea what underlying
IO scheduler is and what are IO scheduler's policies. So it can't keep up
with IO scheduler's policies. Secondly, it might be a low weight group and
tokens might not be available fast enough to release the request.

Read Vs Write
-------------
Writes can overwhelm readers hence second level controller FIFO release
will run into issue here. If there is a single queue maintained then reads
will suffer large latencies. If there separate queues for reads and writes
then it will be hard to decide in what ratio to dispatch reads and writes as
it is IO scheduler's decision to decide when and how much read/write to
dispatch. This is another place where higher level controller will not be in
sync with lower level io scheduler and can change the effective policies of
underlying io scheduler.

The IO schedulers already take care of read-vs-write and already take
care of preventing large writes-starve-reads latencies (or at least,
they're supposed to).

True. Actually this is a limitation of higher level controller. A higher
level controller will most likely implement some of kind of queuing/buffering
mechanism where it will buffer requeuests when it decides to throttle the
group. Now once a fair number read and requests are buffered, and if
controller is ready to dispatch some requests from the group, which
requests/bio should it dispatch? reads first or writes first or reads and
writes in certain ratio?

In what ratio reads and writes are dispatched is the property and decision of
IO scheduler. Now higher level controller will be taking this decision and
change the behavior of underlying io scheduler.


CFQ IO context Issues
---------------------
Buffering at higher layer means submission of bios later with the help of
a worker thread.

Why?

If it's a read, we just block the userspace process.

If it's a delayed write, the IO submission already happens in a kernel thread.

Is it ok to block pdflush on group. Some low weight group might block it
for long time and hence not allow flushing out other pages. Probably that's
the reason pdflush used to check if underlying device is congested or not
and if it is congested, we don't go ahead with submission of request.
With per bdi flusher thread things will change.

I think btrfs also has some threds which don't want to block and if
underlying deivce is congested, it bails out. That's the reason I
implemented per group congestion interface where if a thread does not want
to block, it can check whether the group IO is going in is congested or
not and will it block. So for such threads, probably higher level
controller shall have to implement per group congestion interface so that
threads which don't want to block can check with the controller whether
it has sufficient BW to let it through and not block or may be start
buffering writes in group queue.


If it's a synchronous write, we have to block the userspace caller
anyway.

Async reads might be an issue, dunno.


I think async IO is one of the reason. IIRC, Andrea Righi, implemented the
policy of returning error for async IO if group did not have sufficient
tokens to dispatch the async IO and expected the application to retry
later. I am not sure if that is ok.

So yes, if we are not buffering any of the read requests and either
blocking the caller or returning an error (async IO) than CFQ io context is not
an issue.

This changes the io context information at CFQ layer which
assigns the request to submitting thread. Change of io context info again
leads to issues of idle timer expiry and issue of a process not getting fair
share and reduced throughput.

But we already have that problem with delayed writeback, which is a
huge thing - often it's the majority of IO.


For delayed writes CFQ will not anticipate so increased anticipation timer
expiry is not an issue with writes. But it probably will be an issue with
reads where if higher level controller decides to block next read and
CFQ is anticipating on that read. I am wondering that such kind of issues
must appear with all the higher level device mapper/software raid devices
also. How do they handle it. May be it is more theoritical and in practice
impact is not significant.

Throughput with noop, deadline and AS
---------------------------------------------
I think an higher level controller will result in reduced overall throughput
(as compared to io scheduler based io controller) and more seeks with noop,
deadline and AS.

The reason being, that it is likely that IO with-in a group will be related
and will be relatively close as compared to IO across the groups. For example,
thread pool of kvm-qemu doing IO for virtual machine. In case of higher level
control, IO from various groups will go into a single queue at lower level
controller and it might happen that IO is now interleaved (G1, G2, G1, G3,
G4....) causing more seeks and reduced throughput. (Agreed that merging will
help up to some extent but still....).

Instead, in case of lower level controller, IO scheduler maintains one queue
per group hence there is no interleaving of IO between groups. And if IO is
related with-in group, then we shoud get reduced number/amount of seek and
higher throughput.

Latency can be a concern but that can be controlled by reducing the time
slice length of the queue.

Well maybe, maybe not. If a group is throttled, it isn't submitting
new IO. The unthrottled group is doing the IO submitting and that IO
will have decent locality.

But throttling will kick in ocassionaly. Rest of the time both the groups
will be dispatching bios at the same time. So for most part of it IO
scheduler will probably see IO from both the groups and there will be
small intervals where one group is completely throttled and IO scheduler
is busy dispatching requests only from a single group.


Fairness at logical device level vs at physical device level
------------------------------------------------------------

IO scheduler based controller has the limitation that it works only with the
bottom most devices in the IO stack where IO scheduler is attached.

For example, assume a user has created a logical device lv0 using three
underlying disks sda, sdb and sdc. Also assume there are two tasks T1 and T2
in two groups doing IO on lv0. Also assume that weights of groups are in the
ratio of 2:1 so T1 should get double the BW of T2 on lv0 device.

T1 T2
\ /
lv0
/ | \
sda sdb sdc


Now resource control will take place only on devices sda, sdb and sdc and
not at lv0 level. So if IO from two tasks is relatively uniformly
distributed across the disks then T1 and T2 will see the throughput ratio
in proportion to weight specified. But if IO from T1 and T2 is going to
different disks and there is no contention then at higher level they both
will see same BW.

Here a second level controller can produce better fairness numbers at
logical device but most likely at redued overall throughput of the system,
because it will try to control IO even if there is no contention at phsical
possibly leaving diksks unused in the system.

Hence, question comes that how important it is to control bandwidth at
higher level logical devices also. The actual contention for resources is
at the leaf block device so it probably makes sense to do any kind of
control there and not at the intermediate devices. Secondly probably it
also means better use of available resources.

hm. What will be the effects of this limitation in real-world use?

In some cases user/application will not see the bandwidth ratio between
two groups in same proportion as assigned weights and primary reason for
that will be that this workload did not create enough contention for
physical resources unerneath.

So it all depends on what kind of bandwidth gurantees are we offering. If
we are saying that we provide good fairness numbers at logical devices
irrespective of whether resources are not used optimally, then it will be
irritating for the user.

I think it also might become an issue once we implement max bandwidth
control. We will not be able to define max bandwidth on a logical device
and an application will get more than max bandwidth if it is doing IO to
different underlying devices.

I would say that leaf node control is good for optimal resource usage and
for proportional BW control, but not a good fit for max bandwidth control.


Limited Fairness
----------------
Currently CFQ idles on a sequential reader queue to make sure it gets its
fair share. A second level controller will find it tricky to anticipate.
Either it will not have any anticipation logic and in that case it will not
provide fairness to single readers in a group (as dm-ioband does) or if it
starts anticipating then we should run into these strange situations where
second level controller is anticipating on one queue/group and underlying
IO scheduler might be anticipating on something else.

It depends on the size of the inter-group timeslices. If the amount of
time for which a group is unthrottled is "large" comapred to the
typical anticipation times, this issue fades away.

And those timeslices _should_ be large. Because as you mentioned
above, different groups are probably working different parts of the
disk.

Need of device mapper tools
---------------------------
A device mapper based solution will require creation of a ioband device
on each physical/logical device one wants to control. So it requires usage
of device mapper tools even for the people who are not using device mapper.
At the same time creation of ioband device on each partition in the system to
control the IO can be cumbersome and overwhelming if system has got lots of
disks and partitions with-in.


IMHO, IO scheduler based IO controller is a reasonable approach to solve the
problem of group bandwidth control, and can do hierarchical IO scheduling
more tightly and efficiently.

But I am all ears to alternative approaches and suggestions how doing things
can be done better and will be glad to implement it.

TODO
====
- code cleanups, testing, bug fixing, optimizations, benchmarking etc...
- More testing to make sure there are no regressions in CFQ.

Testing
=======

Environment
==========
A 7200 RPM SATA drive with queue depth of 31. Ext3 filesystem.

That's a bit of a toy.

Yes it is. :-)


Do we have testing results for more enterprisey hardware? Big storage
arrays? SSD? Infiniband? iscsi? nfs? (lol, gotcha)

Not yet. I will try to get hold of some storage arrays and run some tests.



I am mostly
running fio jobs which have been limited to 30 seconds run and then monitored
the throughput and latency.

Test1: Random Reader Vs Random Writers
======================================
Launched a random reader and then increasing number of random writers to see
the effect on random reader BW and max lantecies.

[fio --rw=randwrite --bs=64K --size=2G --runtime=30 --direct=1 --ioengine=libaio --iodepth=4 --numjobs= <1 to 32> ]
[fio --rw=randread --bs=4K --size=2G --runtime=30 --direct=1]

[Vanilla CFQ, No groups]
<--------------random writers--------------------> <------random reader-->
nr Max-bdwidth Min-bdwidth Agg-bdwidth Max-latency Agg-bdwidth Max-latency
1 5737KiB/s 5737KiB/s 5737KiB/s 164K usec 503KiB/s 159K usec
2 2055KiB/s 1984KiB/s 4039KiB/s 1459K usec 150KiB/s 170K usec
4 1238KiB/s 932KiB/s 4419KiB/s 4332K usec 153KiB/s 225K usec
8 1059KiB/s 929KiB/s 7901KiB/s 1260K usec 118KiB/s 377K usec
16 604KiB/s 483KiB/s 8519KiB/s 3081K usec 47KiB/s 756K usec
32 367KiB/s 222KiB/s 9643KiB/s 5940K usec 22KiB/s 923K usec

Created two cgroups group1 and group2 of weights 500 each. Launched increasing
number of random writers in group1 and one random reader in group2 using fio.

[IO controller CFQ; group_idle=8; group1 weight=500; group2 weight=500]
<--------------random writers(group1)-------------> <-random reader(group2)->
nr Max-bdwidth Min-bdwidth Agg-bdwidth Max-latency Agg-bdwidth Max-latency
1 18115KiB/s 18115KiB/s 18115KiB/s 604K usec 345KiB/s 176K usec
2 3752KiB/s 3676KiB/s 7427KiB/s 4367K usec 402KiB/s 187K usec
4 1951KiB/s 1863KiB/s 7642KiB/s 1989K usec 384KiB/s 181K usec
8 755KiB/s 629KiB/s 5683KiB/s 2133K usec 366KiB/s 319K usec
16 418KiB/s 369KiB/s 6276KiB/s 1323K usec 352KiB/s 287K usec
32 236KiB/s 191KiB/s 6518KiB/s 1910K usec 337KiB/s 273K usec

That's a good result.

Also ran the same test with IO controller CFQ in flat mode to see if there
are any major deviations from Vanilla CFQ. Does not look like any.

[IO controller CFQ; No groups ]
<--------------random writers--------------------> <------random reader-->
nr Max-bdwidth Min-bdwidth Agg-bdwidth Max-latency Agg-bdwidth Max-latency
1 5696KiB/s 5696KiB/s 5696KiB/s 259K usec 500KiB/s 194K usec
2 2483KiB/s 2197KiB/s 4680KiB/s 887K usec 150KiB/s 159K usec
4 1471KiB/s 1433KiB/s 5817KiB/s 962K usec 126KiB/s 189K usec
8 691KiB/s 580KiB/s 5159KiB/s 2752K usec 197KiB/s 246K usec
16 781KiB/s 698KiB/s 11892KiB/s 943K usec 61KiB/s 529K usec
32 415KiB/s 324KiB/s 12461KiB/s 4614K usec 17KiB/s 737K usec

Notes:
- With vanilla CFQ, random writers can overwhelm a random reader. Bring down
its throughput and bump up latencies significantly.

Isn't that a CFQ shortcoming which we should address separately? If
so, the comparisons aren't presently valid because we're comparing with
a CFQ which has known, should-be-fixed problems.

I am not sure if it is a CFQ issue. These are synchronous random writes.
These are equally important as random reader. So now CFQ has 33 synchronous
queues to serve. Becuase it does not know about groups, it has no choice but
to serve them in round robin manner. So it does not sound like a CFQ issue.
Just that CFQ can give random reader an advantage if it knows that random
reader is in a different group and that's where IO controller comes in to
picture.


- With IO controller, one can provide isolation to the random reader group and
maintain consitent view of bandwidth and latencies.

Test2: Random Reader Vs Sequential Reader
========================================
Launched a random reader and then increasing number of sequential readers to
see the effect on BW and latencies of random reader.

[fio --rw=read --bs=4K --size=2G --runtime=30 --direct=1 --numjobs= <1 to 16> ]
[fio --rw=randread --bs=4K --size=2G --runtime=30 --direct=1]

[ Vanilla CFQ, No groups ]
<---------------seq readers----------------------> <------random reader-->
nr Max-bdwidth Min-bdwidth Agg-bdwidth Max-latency Agg-bdwidth Max-latency
1 23318KiB/s 23318KiB/s 23318KiB/s 55940 usec 36KiB/s 247K usec
2 14732KiB/s 11406KiB/s 26126KiB/s 142K usec 20KiB/s 446K usec
4 9417KiB/s 5169KiB/s 27338KiB/s 404K usec 10KiB/s 993K usec
8 3360KiB/s 3041KiB/s 25850KiB/s 954K usec 60KiB/s 956K usec
16 1888KiB/s 1457KiB/s 26763KiB/s 1871K usec 28KiB/s 1868K usec

Created two cgroups group1 and group2 of weights 500 each. Launched increasing
number of sequential readers in group1 and one random reader in group2 using
fio.

[IO controller CFQ; group_idle=1; group1 weight=500; group2 weight=500]
<---------------group1---------------------------> <------group2--------->
nr Max-bdwidth Min-bdwidth Agg-bdwidth Max-latency Agg-bdwidth Max-latency
1 13733KiB/s 13733KiB/s 13733KiB/s 247K usec 330KiB/s 154K usec
2 8553KiB/s 4963KiB/s 13514KiB/s 472K usec 322KiB/s 174K usec
4 5045KiB/s 1367KiB/s 13134KiB/s 947K usec 318KiB/s 178K usec
8 1774KiB/s 1420KiB/s 13035KiB/s 1871K usec 323KiB/s 233K usec
16 959KiB/s 518KiB/s 12691KiB/s 3809K usec 324KiB/s 208K usec

Also ran the same test with IO controller CFQ in flat mode to see if there
are any major deviations from Vanilla CFQ. Does not look like any.

[IO controller CFQ; No groups ]
<---------------seq readers----------------------> <------random reader-->
nr Max-bdwidth Min-bdwidth Agg-bdwidth Max-latency Agg-bdwidth Max-latency
1 23028KiB/s 23028KiB/s 23028KiB/s 47460 usec 36KiB/s 253K usec
2 14452KiB/s 11176KiB/s 25628KiB/s 145K usec 20KiB/s 447K usec
4 8815KiB/s 5720KiB/s 27121KiB/s 396K usec 10KiB/s 968K usec
8 3335KiB/s 2827KiB/s 24866KiB/s 960K usec 62KiB/s 955K usec
16 1784KiB/s 1311KiB/s 26537KiB/s 1883K usec 26KiB/s 1866K usec

Notes:
- The BW and latencies of random reader in group 2 seems to be stable and
bounded and does not get impacted much as number of sequential readers
increase in group1. Hence provding good isolation.

- Throughput of sequential readers comes down and latencies go up as half
of disk bandwidth (in terms of time) has been reserved for random reader
group.

Test3: Sequential Reader Vs Sequential Reader
============================================
Created two cgroups group1 and group2 of weights 500 and 1000 respectively.
Launched increasing number of sequential readers in group1 and one sequential
reader in group2 using fio and monitored how bandwidth is being distributed
between two groups.

First 5 columns give stats about job in group1 and last two columns give
stats about job in group2.

<---------------group1---------------------------> <------group2--------->
nr Max-bdwidth Min-bdwidth Agg-bdwidth Max-latency Agg-bdwidth Max-latency
1 8970KiB/s 8970KiB/s 8970KiB/s 230K usec 20681KiB/s 124K usec
2 6783KiB/s 3202KiB/s 9984KiB/s 546K usec 19682KiB/s 139K usec
4 4641KiB/s 1029KiB/s 9280KiB/s 1185K usec 19235KiB/s 172K usec
8 1435KiB/s 1079KiB/s 9926KiB/s 2461K usec 19501KiB/s 153K usec
16 764KiB/s 398KiB/s 9395KiB/s 4986K usec 19367KiB/s 172K usec

Note: group2 is getting double the bandwidth of group1 even in the face
of increasing number of readers in group1.

Test4 (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.

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 buffered 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.

Here's where it all falls to pieces.

For async writeback we just don't care about IO priorities. Because
from the point of view of the userspace task, the write was async! It
occurred at memory bandwidth speed.

It's only when the kernel's dirty memory thresholds start to get
exceeded that we start to care about prioritisation. And at that time,
all dirty memory (within a memcg?) is equal - a high-ioprio dirty page
consumes just as much memory as a low-ioprio dirty page.

So when balance_dirty_pages() hits, what do we want to do?

I suppose that all we can do is to block low-ioprio processes more
agressively at the VFS layer, to reduce the rate at which they're
dirtying memory so as to give high-ioprio processes more of the disk
bandwidth.

But you've gone and implemented all of this stuff at the io-controller
level and not at the VFS level so you're, umm, screwed.

True that's an issue. For async writes we don't create parallel IO paths
from user space to IO scheduler hence it is hard to provide fairness in
all the cases. I think part of the problem is page cache and some
serialization also comes from kjournald.

How about coming up with another cgroup controller for buffered writes or
clubbing it with memory controller as KAMEZAWA Hiroyuki suggested and co-mount
this with io controller? This should help control buffered writes per
cgroup.


Importantly screwed! It's a very common workload pattern, and one
which causes tremendous amounts of IO to be generated very quickly,
traditionally causing bad latency effects all over the place. And we
have no answer to this.

Vanilla CFQ Vs IO Controller CFQ
================================
We have not fundamentally changed CFQ, instead enhanced it to also support
hierarchical io scheduling. In the process invariably there are small changes
here and there as new scenarios come up. Running some tests here and comparing
both the CFQ's to see if there is any major deviation in behavior.

Test1: Sequential Readers
=========================
[fio --rw=read --bs=4K --size=2G --runtime=30 --direct=1 --numjobs=<1 to 16> ]

IO scheduler: Vanilla CFQ

nr Max-bdwidth Min-bdwidth Agg-bdwidth Max-latency
1 35499KiB/s 35499KiB/s 35499KiB/s 19195 usec
2 17089KiB/s 13600KiB/s 30690KiB/s 118K usec
4 9165KiB/s 5421KiB/s 29411KiB/s 380K usec
8 3815KiB/s 3423KiB/s 29312KiB/s 830K usec
16 1911KiB/s 1554KiB/s 28921KiB/s 1756K usec

IO scheduler: IO controller CFQ

nr Max-bdwidth Min-bdwidth Agg-bdwidth Max-latency
1 34494KiB/s 34494KiB/s 34494KiB/s 14482 usec
2 16983KiB/s 13632KiB/s 30616KiB/s 123K usec
4 9237KiB/s 5809KiB/s 29631KiB/s 372K usec
8 3901KiB/s 3505KiB/s 29162KiB/s 822K usec
16 1895KiB/s 1653KiB/s 28945KiB/s 1778K usec

Test2: Sequential Writers
=========================
[fio --rw=write --bs=4K --size=2G --runtime=30 --direct=1 --numjobs=<1 to 16> ]

IO scheduler: Vanilla CFQ

nr Max-bdwidth Min-bdwidth Agg-bdwidth Max-latency
1 22669KiB/s 22669KiB/s 22669KiB/s 401K usec
2 14760KiB/s 7419KiB/s 22179KiB/s 571K usec
4 5862KiB/s 5746KiB/s 23174KiB/s 444K usec
8 3377KiB/s 2199KiB/s 22427KiB/s 1057K usec
16 2229KiB/s 556KiB/s 20601KiB/s 5099K usec

IO scheduler: IO Controller CFQ

nr Max-bdwidth Min-bdwidth Agg-bdwidth Max-latency
1 22911KiB/s 22911KiB/s 22911KiB/s 37319 usec
2 11752KiB/s 11632KiB/s 23383KiB/s 245K usec
4 6663KiB/s 5409KiB/s 23207KiB/s 384K usec
8 3161KiB/s 2460KiB/s 22566KiB/s 935K usec
16 1888KiB/s 795KiB/s 21349KiB/s 3009K usec

Test3: Random Readers
=========================
[fio --rw=randread --bs=4K --size=2G --runtime=30 --direct=1 --numjobs=1 to 16]

IO scheduler: Vanilla CFQ

nr Max-bdwidth Min-bdwidth Agg-bdwidth Max-latency
1 484KiB/s 484KiB/s 484KiB/s 22596 usec
2 229KiB/s 196KiB/s 425KiB/s 51111 usec
4 119KiB/s 73KiB/s 405KiB/s 2344 msec
8 93KiB/s 23KiB/s 399KiB/s 2246 msec
16 38KiB/s 8KiB/s 328KiB/s 3965 msec

IO scheduler: IO Controller CFQ

nr Max-bdwidth Min-bdwidth Agg-bdwidth Max-latency
1 483KiB/s 483KiB/s 483KiB/s 29391 usec
2 229KiB/s 196KiB/s 426KiB/s 51625 usec
4 132KiB/s 88KiB/s 417KiB/s 2313 msec
8 79KiB/s 18KiB/s 389KiB/s 2298 msec
16 43KiB/s 9KiB/s 327KiB/s 3905 msec

Test4: Random Writers
=====================
[fio --rw=randwrite --bs=4K --size=2G --runtime=30 --direct=1 --numjobs=1 to 16]

IO scheduler: Vanilla CFQ

nr Max-bdwidth Min-bdwidth Agg-bdwidth Max-latency
1 14641KiB/s 14641KiB/s 14641KiB/s 93045 usec
2 7896KiB/s 1348KiB/s 9245KiB/s 82778 usec
4 2657KiB/s 265KiB/s 6025KiB/s 216K usec
8 951KiB/s 122KiB/s 3386KiB/s 1148K usec
16 66KiB/s 22KiB/s 829KiB/s 1308 msec

IO scheduler: IO Controller CFQ

nr Max-bdwidth Min-bdwidth Agg-bdwidth Max-latency
1 14454KiB/s 14454KiB/s 14454KiB/s 74623 usec
2 4595KiB/s 4104KiB/s 8699KiB/s 135K usec
4 3113KiB/s 334KiB/s 5782KiB/s 200K usec
8 1146KiB/s 95KiB/s 3832KiB/s 593K usec
16 71KiB/s 29KiB/s 814KiB/s 1457 msec

Notes:
- Does not look like that anything has changed significantly.

Previous versions of the patches were 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
(V7) http://lkml.org/lkml/2009/7/24/253
(V8) http://lkml.org/lkml/2009/8/16/204
(V9) http://lkml.org/lkml/2009/8/28/327

Thanks
Vivek
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Quantcast