Re: IO speed limited by size of IO request (for RBD driver)

["Roger Pau Monné" <roger.pau@citrix.com>]

On 08/05/13 12:32, Steven Haigh wrote:
> On 8/05/2013 6:33 PM, Roger Pau Monné wrote:
>> On 08/05/13 10:20, Steven Haigh wrote:
>>> On 30/04/2013 8:07 PM, Felipe Franciosi wrote:
>>>> I noticed you copied your results from "dd", but I didn't see any conclusions drawn from experiment.
>>>>
>>>> Did I understand it wrong or now you have comparable performance on dom0 and domU when using DIRECT?
>>>>
>>>> domU:
>>>> # dd if=/dev/zero of=output.zero bs=1M count=2048 oflag=direct
>>>> 2048+0 records in
>>>> 2048+0 records out
>>>> 2147483648 bytes (2.1 GB) copied, 25.4705 s, 84.3 MB/s
>>>>
>>>> dom0:
>>>> # dd if=/dev/zero of=output.zero bs=1M count=2048 oflag=direct
>>>> 2048+0 records in
>>>> 2048+0 records out
>>>> 2147483648 bytes (2.1 GB) copied, 24.8914 s, 86.3 MB/s
>>>>
>>>>
>>>> I think that if the performance differs when NOT using DIRECT, the issue must be related to the way your guest is flushing the cache. This must be generating a workload that doesn't perform well on Xen's PV protocol.
>>>
>>> Just wondering if there is any further input on this... While DIRECT
>>> writes are as good as can be expected, NON-DIRECT writes in certain
>>> cases (specifically with a mdadm raid in the Dom0) are affected by about
>>> a 50% loss in throughput...
>>>
>>> The hard part is that this is the default mode of writing!
>>
>> As another test with indirect descriptors, could you change
>> xen_blkif_max_segments in xen-blkfront.c to 128 (it is 32 by default),
>> recompile the DomU kernel and see if that helps?
> 
> Ok, here we go.... compiled as 3.8.0-2 with the above change. 3.8.0-2 is 
> running on both the Dom0 and DomU.
> 
> # dd if=/dev/zero of=output.zero bs=1M count=2048
> 2048+0 records in
> 2048+0 records out
> 2147483648 bytes (2.1 GB) copied, 22.1703 s, 96.9 MB/s
> 
> avg-cpu:  %user   %nice %system %iowait  %steal   %idle
>             0.34    0.00   17.10    0.00    0.23   82.33
> 
> Device:         rrqm/s   wrqm/s     r/s     w/s    rMB/s    wMB/s 
> avgrq-sz avgqu-sz   await  svctm  %util
> sdd             980.97 11936.47   53.11  429.78     4.00    48.77 
> 223.81    12.75   26.10   2.11 101.79
> sdc             872.71 11957.87   45.98  435.67     3.55    49.30 
> 224.71    13.77   28.43   2.11 101.49
> sde             949.26 11981.88   51.30  429.33     3.91    48.90 
> 225.03    21.29   43.91   2.27 109.08
> sdf             915.52 11968.52   48.58  428.88     3.73    48.92 
> 225.84    21.44   44.68   2.27 108.56
> md2               0.00     0.00    0.00 1155.61     0.00    97.51 
> 172.80     0.00    0.00   0.00   0.00
> 
> # dd if=/dev/zero of=output.zero bs=1M count=2048 oflag=direct
> 2048+0 records in
> 2048+0 records out
> 2147483648 bytes (2.1 GB) copied, 25.3708 s, 84.6 MB/s
> 
> avg-cpu:  %user   %nice %system %iowait  %steal   %idle
>             0.11    0.00   13.92    0.00    0.22   85.75
> 
> Device:         rrqm/s   wrqm/s     r/s     w/s    rMB/s    wMB/s 
> avgrq-sz avgqu-sz   await  svctm  %util
> sdd               0.00 13986.08    0.00  263.20     0.00    55.76 
> 433.87     0.43    1.63   1.07  28.27
> sdc             202.10 13741.55    6.52  256.57     0.81    54.77 
> 432.65     0.50    1.88   1.25  32.78
> sde              47.96 11437.57    1.55  261.77     0.19    45.79 
> 357.63     0.80    3.02   1.85  48.60
> sdf            2233.37 11756.13   71.93  191.38     8.99    46.80 
> 433.90     1.49    5.66   3.27  86.15
> md2               0.00     0.00    0.00  731.93     0.00    91.49 
> 256.00     0.00    0.00   0.00   0.00
> 
> Now this is pretty much exactly what I would expect the system to do.... 
> ~96MB/sec buffered, and 85MB/sec direct.

I'm sorry to be such a PITA, but could you also try with 64? If we have
to increase the maximum number of indirect descriptors I would like to
set it to the lowest value that provides good performance to prevent
using too much memory.

> So - it turns out that xen_blkif_max_segments at 32 is a killer in the 
> DomU. Now it makes me wonder what we can do about this in kernels that 
> don't have your series of patches against it? And also about the backend 
> stuff in 3.8.x etc?

There isn't much we can do regarding kernels without indirect
descriptors, there's no easy way to increase the number of segments in a
request.