Re: [PATCH] io-controller: implement per group request allocation limitation

[Vivek Goyal <vgoyal@redhat.com>]

On Mon, Aug 03, 2009 at 10:00:42PM -0400, Munehiro Ikeda wrote:
> Gui Jianfeng wrote, on 07/14/2009 03:45 AM:
>> Munehiro Ikeda wrote:
>>> Vivek Goyal wrote, on 07/13/2009 12:03 PM:
>>>> On Fri, Jul 10, 2009 at 09:56:21AM +0800, Gui Jianfeng wrote:
>>>>> Hi Vivek,
>>>>>
>>>>> This patch exports a cgroup based per group request limits interface.
>>>>> and removes the global one. Now we can use this interface to perform
>>>>> different request allocation limitation for different groups.
>>>>>
>>>> Thanks Gui. Few points come to mind.
>>>>
>>>> - You seem to be making this as per cgroup limit on all devices. I guess
>>>>     that different devices in the system can have different settings of
>>>>     q->nr_requests and hence will probably want different per group limit.
>>>>     So we might have to make it per cgroup per device limit.
>>>  From the viewpoint of implementation, there is a difficulty in my mind to
>>> implement per cgroup per device limit arising from that io_group is
>>> allocated
>>> when associated device is firstly used.  I guess Gui chose per cgroup limit
>>> on all devices approach because of this, right?
>>
>>    Yes, I choose this solution from the simplicity point of view, the code will
>>    get complicated if choosing per cgroup per device limit. But it seems per
>>    cgroup per device limits is more reasonable.
>>
>>>
>>>> - There does not seem to be any checks for making sure that children
>>>>     cgroups don't have more request descriptors allocated than parent
>>>> group.
>>>>
>>>> - I am re-thinking that what's the advantage of configuring request
>>>>     descriptors also through cgroups. It does bring in additional
>>>> complexity
>>>>     with it and it should justfiy the advantages. Can you think of some?
>>>>
>>>>     Until and unless we can come up with some significant advantages, I
>>>> will
>>>>     prefer to continue to use per group limit through q->nr_group_requests
>>>>     interface instead of cgroup. Once things stablize, we can revisit
>>>> it and
>>>>     see how this interface can be improved.
>>> I agree.  I will try to clarify if per group per device limitation is
>>> needed
>>> or not (or, if it has the advantage beyond the complexity) through some
>>> tests.
>>
>>    Great, hope to hear you soon.
>
> Sorry for so late.  I tried it, and write the result and my opinion  
> below...
>
>

Hi Ikeda,

Nice analysis. Few questions/comments inline...


> Scenario
> ====================
>
> The possible scenario where per-cgroup nr_requests limitation is  
> beneficial in my mind is that:
>
> - Process P1 in cgroup "g1" is running with submitting many requests
>    to a device.  The number of the requests in the device queue is
>    almost nr_requests for the device.
>
> - After a while, process P2 in cgroup "g2" starts running.  P2 also
>    tries to submit requests as many as P1.
>
> - Assuming that user wants P2 to grab bandwidth as soon as possible
>    and keep it certain level.
>
> In this scenario, I predicted the bandwidth behavior of P2 along with  
> tuning global nr_group_requests like below.
>
> - If having nr_group_requests almost same as nr_requests, P1 can
>    allocate requests up to nr_requests and there is no room for P2 at
>    the beginning of its running.  As a result of it, P2 has to wait
>    for a while till P1's requests are completed and rising of
>    bandwidth is delayed.
>
> - If having nr_group_requests fewer to restrict requests from P1 and
>    make room for P2, the bandwidth of P2 may be lower than the case
>    that P1 can allocate more requests.
>
> If the prediction is correct and per-cgroup nr_requests limitation can  
> make the situation better, per-cgroup nr_requests is supposed to be  
> beneficial.
>
>
> Verification Conditions
> ========================
>
> - Kernel:
>    2.6.31-rc1
>    + Patches from Vivek on Jul 2, 2009
>      (IO scheduler based IO controller V6)
>
> https://lists.linux-foundation.org/pipermail/containers/2009-July/018948.html
>    + Patches from Gui Jianfeng on Jul 7, 2009 (Bug fixes)
>
> https://lists.linux-foundation.org/pipermail/containers/2009-July/019086.html
>
> https://lists.linux-foundation.org/pipermail/containers/2009-July/019087.html
>    + Patch from Gui Jianfeng on Jul 9, 2009 (per-cgroup requests limit)
>
> https://lists.linux-foundation.org/pipermail/containers/2009-July/019123.html
>    + Patch from me on Jul 16, 2009 (Bug fix)
>
> https://lists.linux-foundation.org/pipermail/containers/2009-July/019286.html
>    + 2 local bug fix patches
>        (Not posted yet, I'm posting them in following mails)
>
> - All results are measured under nr_requests=500.
>
> - Used fio to make I/O.  Job file is like below.  Used libaio and
>    direct-I/O and tuned iodepth to make rl->count[1] approx 500 always.
>
> ----- fio job file : from here -----
>
> [global]
> size=128m
> directory=/mnt/b1
>
> runtime=30
> time_based
>
> write_bw_log
> bwavgtime=200
>
> rw=randread
> direct=1
> ioengine=libaio
> iodepth=500
>
> [g1]
> exec_prerun=./pre.sh /mnt/cgroups/g1
> exec_postrun=./log.sh /mnt/cgroups/g1 sdb "_post"
>
> [g2]
> startdelay=10
> exec_prerun=./pre.sh /mnt/cgroups/g2
> exec_postrun=./log.sh /mnt/cgroups/g2 sdb "_post"
>
> ----- fio job file : till here -----
>
> Note:
> pre.sh and log.sh used in exec_{pre|post}run are to assign processes to 
> expected cgroups and record the conditions.  Let me omit the detail of 
> them because they are not fundamental part of this verification.
>
>
> Results
> ====================
>
> Bandwidth of g2 (=P2) was measured under some conditions.  Conditions  
> and bandwidth logs are shown below.
> Bandwidth logs are shown only the beginning part (from starting of P2 to 
> 3000[ms] after aprox.) because the full logs are so long.  Average  
> bandwidth from the beginning of log to ~10[sec] is also calculated.
>
> Note1:
> fio seems to log bandwidth only when actual data transfer occurs  
> (correct me if it's not true).  This means that there is no line with  
> BW=0.  In there is no data transfer, the time-stamp are simply skipped  
> to record.
>
> Note2:
> Graph picture of the bandwidth logs is attached.
>    Result(1): orange
>    Result(2): green
>    Result(3): brown
>    Result(4): black
>
>
> ---------- Result (1) ----------
>
> * Both of g1 and g2 have nr_group_requests=500
>
> < Conditions >
> nr_requests = 500
> g1/
>    io.nr_group_requests = 500
>    io.weight = 500
>    io.ioprio_class = 2
> g2/
>    io.nr_group_requests = 500
>    io.weight = 500
>    io.ioprio_class = 2
>
> < Bandwidth log of g2 >
> t [ms]	bw[KiB/s]
> 969	4
> 1170	1126
> 1374	1084
> 1576	876
> 1776	901
> 1980	1069
> 2191	1087
> 2400	1117
> 2612	1087
> 2822	1136
> ...
>
> < Average bandwidth >
> 1063 [KiB/s]
> (969~9979[ms])
>
>
> ---------- Result (2) ----------
>
> * Both of g1 and g2 have nr_group_requests=100
>
> < Conditions >
> nr_requests = 500
> g1/
>    io.nr_group_requests = 100
>    io.weight = 500
>    io.ioprio_class = 2
> g2/
>    io.nr_group_requests = 100
>    io.weight = 500
>    io.ioprio_class = 2
>
> < Bandwidth log of g2 >
> t [ms]	bw[KiB/s]
> 1498	2
> 1733	892
> 2096	722
> 2311	1224
> 2534	1180
> 2753	1197
> 2988	1137
> ...
>
> < Average bandwidth >
> 998 [KiB/s]
> (1498~9898[ms])
>
>
> ---------- Result (3) ----------
>
> * To set different nr_group_requests on g1 and g2
>
> < Conditions >
> nr_requests = 500
> g1/
>    io.nr_group_requests = 100
>    io.weight = 500
>    io.ioprio_class = 2
> g2/
>    io.nr_group_requests = 500
>    io.weight = 500
>    io.ioprio_class = 2
>
> < Bandwidth log of g2 >
> t [ms]	bw[KiB/s]
> 244	839
> 451	1133
> 659	964
> 877	1038
> 1088	1125
> 1294	979
> 1501	1068
> 1708	934
> 1916	1048
> 2117	1126
> 2328	1111
> 2533	1118
> 2758	1206
> 2969	990
> ...
>
> < Average bandwidth >
> 1048 [KiB/s]
> (244~9906[ms])
>
>
> ---------- Result (4) ----------
>
> * To make g2/io.ioprio_class as RT
>
> < Conditions >
> nr_requests = 500
> g1/
>    io.nr_group_requests = 500
>    io.weight = 500
>    io.ioprio_class = 2
> g2/
>    io.nr_group_requests = 500
>    io.weight = 500
>    io.ioprio_class = 1
>
> < Bandwidth log of g2 >
> t [ms]	bw[KiB/s]
> 476	8
> 677	2211
> 878	2221
> 1080	2486
> 1281	2241
> 1481	2109
> 1681	2334
> 1882	2129
> 2082	2211
> 2283	1915
> 2489	1778
> 2690	1915
> 2891	1997
> ...
>
> < Average bandwidth >
> 2132[KiB/s]
> (476~9954[ms])
>
>
> Consideration and Conclusion
> =============================
>
>  From result(1), it is observed that it takes 1000~1200[ms] to rise P2 
> bandwidth.  In result(2), where both of g1 and g2 have  
> nr_group_requests=100, the delay gets longer as 1800~2000[ms].

Result (2) is surprising in terms of delay. Queue limit is 500 and per
group limit is 100. That means that group g2 should have got request
descriptor allocated as soon as it sent a request to device. That should
also mean that it should start getting serviced soon and BW available
should rise soon. 

I am not sure why reverse is happening. Sounds like a bug somewhere..
 
>  In  
> addition to it, the average bandwidth becomes ~5% lower than result(1).  
> This is supposed that P2 couldn't allocate enough requests.
> Then, result(3) shows that bandwidth of P2 can rise quickly (~300[ms])  
> if nr_group_requests can be set per-cgroup.  Result(4) shows that the  
> delay can be shortened by setting g2 as RT class, however, the delay is 
> still longer than result(3).
>
> I think it is confirmed that "per-cgroup nr_requests limitation is  
> useful in a certain situation".

I am still not clear how per cgroup request descriptor limit is useful.
Having higher number of descriptor reserved for a cgroup probably should
help a bit on queuing hardware where more requests from same queue can be
dispatched at one go. May be it can also help a bit in more merging in 
certain scenarios and gives higher throughput.

I think key question here is how relevant the number of request
descriptors is when it comes to fairness for the group. I have not
experimented but probably a few request descriptors (25-30) probably
should be enough to ensure group gets fair amount of disk. That's a 
different thing that throughput of group might suffer and bandwith
allocation among task with-in group will also have adverse effect.

>  Beyond that, the discussion topic is  
> the benefit pointed out above is eligible for the complication of the  
> implementation.  IMHO, I don't think the implementation of per-cgroup  
> request limitation is too complicated to accept.  On the other hand I  
> guess it suddenly gets complicated if we try to implement further more, 
> especially hierarchical support.  It is also true that I have a feeling 
> that implementation without per-device limitation and hierarchical 
> support is like "unfinished work".
>

IMHO, configuring request descriptors per cgroup can be a future TODO item
depending on how useful people consider it. For the time being, it would 
be better to keep things simple and code small.

In fact currently I am trying to replace BFQ with a CFS style scheduler in
io controller patches to reduce the code size as well as its complexity.

Thanks
Vivek

> So, my opinion so far is that, per-cgroup nr_requests limitation should 
> be merged only if hierarchical support is concluded "unnecessary" for it. 
>  If merging it tempts hierarchical support, it shouldn't be.
> How about your opinion, all?
>
> My considerations or verification method might be wrong.  Please correct 
> them if any.  And if you have any other idea of scenario to verify the 
> effect of per-cgroup nr_requests limitation, please let me know.  I'll 
> try it.
>
>
>
> -- 
> IKEDA, Munehiro
>    NEC Corporation of America
>      m-ikeda@ds.jp.nec.com
>
>