Re: Xen disk write slowness in kernel 3.8.x

[Steven Haigh <netwiz@crc.id.au>]

On 18/04/2013, at 23:10, Felipe Franciosi <felipe.franciosi@citrix.com> 
wrote:

> Hi Steven,
>
> I'm a bit surprised with the results of your "dd" tests using bs=11M.
> I'd expect to see a constant avgrq-sz of 88 (44K). You seem to be experiencing a varying number near 66.
>
> I noticed your "dd" was done on top of a mount point. Could you repeat that test on top of the raw /dev/xvd* device?

Sadly not. The file system takes up the whole RAID6 array and is XFS. As 
previously stated, the device on Dom0 (/dev/md2) is passed as xvdb to 
the DomU.

XFS Info from the DomU:
$ xfs_info /dev/xvdb
meta-data=/dev/xvdb              isize=256    agcount=32, 
agsize=30521632 blks
          =                       sectsz=4096  attr=2, projid32bit=0
data     =                       bsize=4096   blocks=976691648, imaxpct=5
          =                       sunit=32     swidth=64 blks
naming   =version 2              bsize=4096   ascii-ci=0
log      =internal               bsize=4096   blocks=476900, version=2
          =                       sectsz=4096  sunit=1 blks, lazy-count=1
realtime =none                   extsz=4096   blocks=0, rtextents=0

> And yes, oprofile should be ran on dom0. If you provide the kernel symbols for your domU, it will be able to give you detailed information on that as well.

Hmmm - not exactly sure how to do this. Any pointers to where I can read 
up on it?

> I'm afraid I don't have RPMs for Roger's kernel tree. It shouldn't be too hard to clone and build it, though. Let me know if you have problems with that.

I'll try to have a look at this over the weekend.

--
Steven Haigh

Email: netwiz@crc.id.au
Web: https://www.crc.id.au
Phone: (03) 9001 6090 - 0412 935 897
Fax: (03) 8338 0299

> -----Original Message-----
> From: Steven Haigh [mailto:netwiz@crc.id.au]
> Sent: 17 April 2013 08:53
> To: Felipe Franciosi
> Cc: xen-devel@lists.xen.org
> Subject: Re: Xen disk write slowness in kernel 3.8.x
>
> On 04/17/2013 12:39 AM, Felipe Franciosi wrote:
>> Hi Steven, sorry for the delay (doing some training followed by holidays).
>>
>> Using "direct" from the guest does speed things up. I'm not entirely sure why. Especially since this is also true (at least with the guests I tested) in XCP/XenServer where the datapath includes a blktap2/tapdisk model where the virtual disks are always opened with O_DIRECT and the only valid operations are READ or WRITE (i.e. there is no special handling for guest operations such as barriers, etc).
>>
>> Have a look at http://support.citrix.com/article/CTX136861 for an understanding of how things work in XCP 1.6 / XenServer 6.1.0.
>>
>> My guess is that "direct" allows you to control the exact size of the requests (creating a consistent data flow). If you omit it and shove all your requests through the guest's buffer cache, what you get in the backend will vary largely in terms of request sizes. This will particularly affect the blkback/blkfront ring utilisation. This hypothesis can hopefully be supported by running iostat in dom0 (e.g. iostat -xm 1) and observing the request sizes. Let me know what you see in your environment. If anyone else in the list have other ideas on this I'd like to hear them as this has been puzzling me too.
>>
>> Another thing you can try is to use "bs=11M" along i/oflag=direct. The "11M" magic number comes from the ring size: with single-page rings, you can have up to 32 requests of 44KiB each in the ring.
>> 32 * 44KiB = 1408 KiB
>> If you try that and observe iostat in dom0, you'll see that the average request size often drops below 44KiB (88 sectors). To keep it consistent I decided to use multiples of that number. The first "whole" number in the order of mebibytes you'll find is 8 rings: 32 * 44 * 1024 * 8 = 11 M.
>
> # dd if=/dev/zero of=/mnt/fileshare/output.zero bs=11M count=300
> 300+0 records in
> 300+0 records out
> 3460300800 bytes (3.5 GB) copied, 73.2253 s, 47.3 MB/s
>
> Dom0 iostat -xm 1:
> avg-cpu:  %user   %nice %system %iowait  %steal   %idle
>            0.00    0.00    9.30    0.00    1.16   89.53
>
> Device:         rrqm/s   wrqm/s     r/s     w/s    rMB/s    wMB/s
> avgrq-sz avgqu-sz   await  svctm  %util
> sde            1267.44  6526.74  100.00  545.35     5.31    27.24
> 103.28    10.28   14.86   1.75 113.02
> sdc            1316.28  6348.84  105.81  546.51     5.69    26.94
> 102.45     4.56    7.02   1.37  89.30
> sdd            1139.53  6269.77  110.47  544.19     5.11    26.85
> 99.99     5.34    8.30   1.35  88.37
> sdf            1286.05  6436.05   96.51  548.84     5.50    27.29
> 104.06     3.52    5.47   1.34  86.51
> md2               0.00     0.00    0.00 1572.09     0.00    51.11
> 66.59     0.00    0.00   0.00   0.00
>
> # dd if=/dev/zero of=/mnt/fileshare/output.zero bs=11M count=300 oflag=direct
> 300+0 records in
> 300+0 records out
> 3460300800 bytes (3.5 GB) copied, 67.5844 s, 51.2 MB/s
>
> avg-cpu:  %user   %nice %system %iowait  %steal   %idle
>            1.14    0.00   12.50    0.00    0.00   86.36
>
> Device:         rrqm/s   wrqm/s     r/s     w/s    rMB/s    wMB/s
> avgrq-sz avgqu-sz   await  svctm  %util
> sde            1189.77  7005.68   89.77  478.41     5.00    29.35
> 123.81     5.62    9.90   1.72  97.50
> sdc            1040.91  6852.27   77.27  480.68     4.37    28.76
> 121.60     2.84    5.09   1.11  61.93
> sdd             911.36  7057.95   67.05  477.27     3.66    29.46
> 124.63     4.31    7.51   1.51  82.39
> sdf            1185.23  7227.27   72.73  495.45     4.90    30.28
> 126.80     7.79   13.73   1.82 103.64
> md2               0.00     0.00    0.00 1727.27     0.00    56.54
> 67.04     0.00    0.00   0.00   0.00
>
> It seems that using a block size on dd actually reduces the total write speed.
>
> Compare this to using 1M block sizes:
>
> # dd if=/dev/zero of=/mnt/fileshare/output.zero bs=1M count=3500 oflag=direct
> 3500+0 records in
> 3500+0 records out
> 3670016000 bytes (3.7 GB) copied, 27.6453 s, 133 MB/s
>
> avg-cpu:  %user   %nice %system %iowait  %steal   %idle
>            0.00    0.00   27.78    0.00    0.00   72.22
>
> Device:         rrqm/s   wrqm/s     r/s     w/s    rMB/s    wMB/s
> avgrq-sz avgqu-sz   await  svctm  %util
> sde               0.00 19040.00    0.00  303.33     0.00    75.83
> 512.00     1.28    4.22   2.81  85.22
> sdc               0.00 19040.00    0.00  302.22     0.00    75.56
> 512.00     0.50    1.67   1.10  33.33
> sdd               0.00 19040.00    0.00  302.22     0.00    75.56
> 512.00     0.50    1.66   1.10  33.33
> sdf               0.00 19040.00    0.00  303.33     0.00    75.83
> 512.00     1.28    4.22   2.81  85.22
> md2               0.00     0.00    0.00 4684.44     0.00   151.11
> 66.06     0.00    0.00   0.00   0.00
>
> # dd if=/dev/zero of=/mnt/fileshare/output.zero bs=1M count=3500
> 3500+0 records in
> 3500+0 records out
> 3670016000 bytes (3.7 GB) copied, 81.3315 s, 45.1 MB/s
>
> avg-cpu:  %user   %nice %system %iowait  %steal   %idle
>            0.00    0.00   10.47    0.00    0.00   89.53
>
> Device:         rrqm/s   wrqm/s     r/s     w/s    rMB/s    wMB/s
> avgrq-sz avgqu-sz   await  svctm  %util
> sde            1116.28  5995.35  101.16  555.81     5.02    25.84
> 96.20     5.99    9.31   1.52 100.00
> sdc            1209.30  6019.77  111.63  536.05     5.19    25.07
> 95.70     5.66    8.58   1.45  93.95
> sdd            1236.05  5967.44   97.67  546.51     5.08    25.54
> 97.34     8.16   12.67   1.60 103.02
> sdf            1203.49  5939.53   97.67  554.65     5.12    25.63
> 96.54     4.76    7.48   1.44  93.95
> md2               0.00     0.00    0.00 1486.05     0.00    48.31
> 66.59     0.00    0.00   0.00   0.00
>
> Out of interest, trying 40k block sizes seems to show roughly the same results as the 11M tests... Maybe this is not related - but I'm not quite sure.
>
>> Verify if that gives you more consistency in your results.
>>
>> Last but not least, you should definitely try two other things:
>> 1) oprofile.
>> 2) Roger's kernel with persistent grants and indirect I/O reqs (see below).
>>
>> Using oprofile you should be able to get more directions/ideas in the sense of where your throughput is getting killed.
>
> I'll have to look at how to use oprofile to give you useful information from it... I assume you mean running this on the Dom0?
>
>> Finally, high throughput is usually achieved with large bandwidth and low latency. Persistent granting should help a lot with reducing latency. The indirect I/O operations that have recently been discussed in this list should help you to achieve larger bandwidth. This should be particularly beneficial in arrays such as yours, where the stripe size is larger than 88K.
>>
>> Link to Roger's blog post and kernel code:
>> http://blog.xen.org/index.php/2012/11/23/improving-block-protocol-scal
>> ability-with-persistent-grants/#more-5711
>> http://xenbits.xen.org/gitweb/?p=people/royger/linux.git;a=summary
>> (xen-block-indirect branch)
>
> I'll do some research into this - if someone has a binary RPM package already built from this, I might be able to get results to you much sooner...
>
>> I am very interested in your results, please share them.
>>
>> Cheers,
>> Felipe
>>
>>
>> -----Original Message-----
>> From: Steven Haigh [mailto:netwiz@crc.id.au]
>> Sent: 09 April 2013 07:54
>> To: Felipe Franciosi
>> Cc: xen-devel@lists.xen.org
>> Subject: Re: Xen disk write slowness in kernel 3.8.x
>>
>> Resending this - hopefully for comment...
>>
>> On 04/04/13 01:30, Felipe Franciosi wrote:
>>> I think you should use i/oflag=direct on your "dd"s to bypass the buffer cache in the guest. Might provide more consistent results.
>>
>> Ah - you are correct. I usually don't forget that. This has however discovered something interesting...
>>
>> Entirely from the DomU:
>> # dd if=/dev/zero of=/mnt/fileshare/output.zero bs=1M count=4096
>> oflag=direct
>> 4096+0 records in
>> 4096+0 records out
>> 4294967296 bytes (4.3 GB) copied, 45.2753 s, 94.9 MB/s # rm
>> /mnt/fileshare/output.zero # dd if=/dev/zero
>> of=/mnt/fileshare/output.zero bs=1M count=4096
>> 4096+0 records in
>> 4096+0 records out
>> 4294967296 bytes (4.3 GB) copied, 80.5808 s, 53.3 MB/s
>>
>> I did this 3-4 times with the same trend...
>>
>> # dd if=/dev/zero of=/mnt/fileshare/output.zero bs=1M count=4096
>> oflag=direct
>> 4096+0 records in
>> 4096+0 records out
>> 4294967296 bytes (4.3 GB) copied, 40.693 s, 106 MB/s # rm
>> /mnt/fileshare/output.zero # sync # dd if=/dev/zero
>> of=/mnt/fileshare/output.zero bs=1M count=4096
>> 4096+0 records in
>> 4096+0 records out
>> 4294967296 bytes (4.3 GB) copied, 82.1511 s, 52.3 MB/s
>>
>> So, when the DomU is using write buffers, write speed is halved. This I don't understand.
>>
>> I haven't tweaked the array at all yet (as I've just been experimenting with this problem) - so with a bit of time, I can probably even increase its write speed some more - but without this speed decrease solved, it isn't really worth it yet...
>>
>>> The result of oprofile is intriguing. Did you have a chance to try Roger's persistent grant implementation?
>>>
>>> He mentions it on his 08/03 e-mail. It's here:
>>> git://xenbits.xen.org/people/royger/linux.git xen-block-indirect
>>
>> I can't say I have - Although I package Xen for EL6, we're getting to the edge of my knowledge here.
>>
>>>
>>> Felipe
>>>
>>> -----Original Message-----
>>> From: Steven Haigh [mailto:netwiz@crc.id.au]
>>> Sent: 03 April 2013 14:27
>>> To: xen-devel@lists.xen.org
>>> Cc: Felipe Franciosi
>>> Subject: Re: Xen disk write slowness in kernel 3.8.x
>>>
>>> On 03/04/13 23:29, Felipe Franciosi wrote:
>>>> Do you know the size of your requests?
>>>> You used "iostat -m". Perhaps "iostat -xm" could be more meaningful as it will tell you the average request size in sectors.
>>>
>>> Good call. I started just a sequential write with dd from /dev/zero on the DomU:
>>> # dd if=/dev/zero of=/mnt/fileshare/output.zero bs=1M count=4096
>>> 4096+0 records in
>>> 4096+0 records out
>>> 4294967296 bytes (4.3 GB) copied, 91.269 s, 47.1 MB/s
>>>
>>> iostat -xm shows:
>>> avg-cpu:  %user   %nice %system %iowait  %steal   %idle
>>>              0.11    0.00    2.71    0.23    0.56   96.38
>>>
>>> Device:         rrqm/s   wrqm/s     r/s     w/s    rMB/s    wMB/s
>>> avgrq-sz avgqu-sz   await  svctm  %util
>>> sdc            1208.40  5337.80  150.00  645.20     5.35    23.39
>>> 74.00     6.47    8.15   0.91  72.72
>>> sdf            1199.80  5350.80  148.00  643.40     5.30    23.34
>>> 74.12     6.88    8.68   0.96  76.06
>>> sdd            1203.40  5304.60  148.80  638.60     5.28    23.21
>>> 74.11     5.42    6.88   0.78  61.38
>>> sde            1213.80  5382.40  148.40  652.40     5.37    23.59
>>> 74.08     6.40    8.00   0.94  75.20
>>> md3               0.00     0.00    1.60 1312.20     0.05    42.68
>>> 66.60     0.00    0.00   0.00   0.00
>>>
>>> Shutting down the DomU, mounting /dev/md3 in and doing this directly from the Dom0 shows:
>>> # dd if=/dev/zero of=/mnt/fileshare/output.zero bs=1M count=4096
>>> 4096+0 records in
>>> 4096+0 records out
>>> 4294967296 bytes (4.3 GB) copied, 44.2801 s, 97.0 MB/s
>>>
>>> details from iostat -xm:
>>> avg-cpu:  %user   %nice %system %iowait  %steal   %idle
>>>              0.00    0.00    2.16    0.00    0.57   97.27
>>>
>>> Device:         rrqm/s   wrqm/s     r/s     w/s    rMB/s    wMB/s
>>> avgrq-sz avgqu-sz   await  svctm  %util
>>> sdc            1175.40  5220.60  143.40  633.60     5.13    22.85
>>> 73.76     6.01    7.73   0.87  67.60
>>> sdf            1175.20  5154.00  147.40  624.40     5.14    22.55
>>> 73.46     6.87    8.88   0.97  74.90
>>> sdd            1183.40  5133.80  145.40  625.20     5.19    22.50
>>> 73.60     5.19    6.73   0.77  59.60
>>> sde            1176.40  5229.60  146.00  637.00     5.16    22.99
>>> 73.62     7.39    9.51   0.99  77.90
>>> md3               0.00     0.00    0.60 1277.40     0.02    41.56
>>> 66.63     0.00    0.00   0.00   0.00
>>>
>>> This about ties in with what bonnie++ gets as the write speed:
>>> Version  1.96       ------Sequential Output------ --Sequential Input-
>>> --Random-
>>> Concurrency   1     -Per Chr- --Block-- -Rewrite- -Per Chr- --Block--
>>> --Seeks--
>>> Machine        Size K/sec %CP K/sec %CP K/sec %CP K/sec %CP K/sec %CP
>>> /sec %CP
>>> xenhost.lan.crc. 2G   635  91 97526  15 67864  16   952  97 295833  35
>>> 409.6  10
>>>
>>> bonnie++ maxes out at about the same as dd in the DomU.
>>>
>>>> I still didn't understand your environment. I think this is the first time you mention NFS/SMB.
>>>
>>> /dev/md3 is passed through as a phy device to a DomU. This DomU runs both samba and NFS shares. It doesn't seem to matter what method is the source of the writes to the drive (dd, bonnie++, sending a file via NFS or Samba), the max write speed seems to be bottlenecking at ~50Mb/sec - which I believe is purely the write speed to the array via the DomU.
>>>
>>>> Weren't you testing copies from /dev/zero to your device?
>>>
>>> I have tested many ways - this being just one of them.
>>>
>>>> I'd also recommend you run "oprofile" to see if there are any time sinks that we are overlooking. You can find out how to use it here:
>>>> http://wiki.xen.org/wiki/Xen_Profiling:_oprofile_and_perf
>>>
>>> I ran 'perf top' on the Dom0, then started a write (dd if=/dev/zero of=/mnt/fileshare/output.zero bs=1M count=4096) on the DomU. This is what I'd see as the relevant output:
>>>
>>> Samples: 574K of event 'cpu-clock', Event count (approx.): 60028
>>>    88.37%  [kernel]             [k] xen_hypercall_sched_op
>>>     4.73%  [kernel]             [k] xen_hypercall_xen_version
>>>     1.00%  [kernel]             [k] xen_hypercall_grant_table_op
>>>     0.99%  [raid456]            [k] handle_stripe
>>>     0.77%  [raid6_pq]           [k] raid6_sse24_gen_syndrome
>>>     0.66%  [raid456]            [k] ops_run_io
>>>     0.57%  [kernel]             [k] memcpy
>>>     0.21%  [kernel]             [k] xen_restore_fl_direct
>>>     0.18%  [raid456]            [k] raid5_end_write_request
>>>     0.18%  [raid456]            [k] __raid_run_ops
>>>     0.14%  [kernel]             [k] xen_hypercall_event_channel_op
>>>     0.11%  [kernel]             [k] get_phys_to_machine
>>>     0.09%  [raid456]            [k] schedule_reconstruction
>>>
>>> Now repeated, with the same on the DomU:
>>> Samples: 300K of event 'cpu-clock', Event count (approx.): 84845
>>>
>>>    97.63%  [kernel]            [k] hypercall_page
>>>     0.44%  [kernel]            [k] copy_user_generic_string
>>>     0.41%  [kernel]            [k] __clear_user
>>>     0.08%  [kernel]            [k] __wake_up_bit
>>>     0.07%  [kernel]            [k] xen_restore_fl_direct
>>>     0.07%  [kernel]            [k] __mem_cgroup_commit_charge
>>>
>>> Now, interestingly enough... The root drive (xvda) of the DomU in question is on a separate RAID1. This is passed as an LV.
>>>
>>> # dd if=/dev/zero of=/output.bin bs=1M count=4096
>>> 4096+0 records in
>>> 4096+0 records out
>>> 4294967296 bytes (4.3 GB) copied, 69.7036 s, 61.6 MB/s
>>>
>>> Then from the Dom0 to the same RAID1:
>>> # dd if=/dev/zero of=output.zero bs=1M count=4096
>>> 4096+0 records in
>>> 4096+0 records out
>>> 4294967296 bytes (4.3 GB) copied, 48.4407 s, 88.7 MB/s
>>>
>>> I don't really know what else I can do to try and see where the
>>> slowdown is here - I am open to suggestions though :)
>>>
>>> --
>>> Steven Haigh
>>>
>>> Email: netwiz@crc.id.au
>>> Web: https://www.crc.id.au
>>> Phone: (03) 9001 6090 - 0412 935 897
>>> Fax: (03) 8338 0299
>>
>>
>> --
>> Steven Haigh
>>
>> Email: netwiz@crc.id.au
>> Web: https://www.crc.id.au
>> Phone: (03) 9001 6090 - 0412 935 897
>> Fax: (03) 8338 0299
>
>
> --
> Steven Haigh
>
> Email: netwiz@crc.id.au
> Web: https://www.crc.id.au
> Phone: (03) 9001 6090 - 0412 935 897
> Fax: (03) 8338 0299