Re: [RESEND] [PATCH] readahead:add blk_run_backing_dev

[Vladislav Bolkhovitin <vst@vlnb.net>]

Ronald Moesbergen, on 07/01/2009 05:07 PM wrote:
> 2009/6/30 Vladislav Bolkhovitin <vst@vlnb.net>:
>> Wu Fengguang, on 06/30/2009 05:04 AM wrote:
>>> On Mon, Jun 29, 2009 at 11:37:41PM +0800, Vladislav Bolkhovitin wrote:
>>>> Wu Fengguang, on 06/29/2009 07:01 PM wrote:
>>>>> On Mon, Jun 29, 2009 at 10:21:24PM +0800, Wu Fengguang wrote:
>>>>>> On Mon, Jun 29, 2009 at 10:00:20PM +0800, Ronald Moesbergen wrote:
>>>>>>> ... tests ...
>>>>>>>
>>>>>>>> We started with 2.6.29, so why not complete with it (to save
>>>>>>>> additional
>>>>>>>> Ronald's effort to move on 2.6.30)?
>>>>>>>>
>>>>>>>>>> 2. Default vanilla 2.6.29 kernel, 512 KB read-ahead, the rest is
>>>>>>>>>> default
>>>>>>>>> How about 2MB RAID readahead size? That transforms into about 512KB
>>>>>>>>> per-disk readahead size.
>>>>>>>> OK. Ronald, can you 4 more test cases, please:
>>>>>>>>
>>>>>>>> 7. Default vanilla 2.6.29 kernel, 2MB read-ahead, the rest is default
>>>>>>>>
>>>>>>>> 8. Default vanilla 2.6.29 kernel, 2MB read-ahead, 64 KB
>>>>>>>> max_sectors_kb, the rest is default
>>>>>>>>
>>>>>>>> 9. Patched by the Fengguang's patch vanilla 2.6.29 kernel, 2MB
>>>>>>>> read-ahead, the rest is default
>>>>>>>>
>>>>>>>> 10. Patched by the Fengguang's patch vanilla 2.6.29 kernel, 2MB
>>>>>>>> read-ahead, 64 KB max_sectors_kb, the rest is default
>>>>>>> The results:
>>>>>> I made a blindless average:
>>>>>>
>>>>>> N       MB/s          IOPS      case
>>>>>>
>>>>>> 0      114.859       984.148    Unpatched, 128KB readahead, 512
>>>>>> max_sectors_kb
>>>>>> 1      122.960       981.213    Unpatched, 512KB readahead, 512
>>>>>> max_sectors_kb
>>>>>> 2      120.709       985.111    Unpatched, 2MB readahead, 512
>>>>>> max_sectors_kb
>>>>>> 3      158.732      1004.714    Unpatched, 512KB readahead, 64
>>>>>> max_sectors_kb
>>>>>> 4      159.237       979.659    Unpatched, 2MB readahead, 64
>>>>>> max_sectors_kb
>>>>>>
>>>>>> 5      114.583       982.998    Patched, 128KB readahead, 512
>>>>>> max_sectors_kb
>>>>>> 6      124.902       987.523    Patched, 512KB readahead, 512
>>>>>> max_sectors_kb
>>>>>> 7      127.373       984.848    Patched, 2MB readahead, 512
>>>>>> max_sectors_kb
>>>>>> 8      161.218       986.698    Patched, 512KB readahead, 64
>>>>>> max_sectors_kb
>>>>>> 9      163.908       574.651    Patched, 2MB readahead, 64
>>>>>> max_sectors_kb
>>>>>>
>>>>>> So before/after patch:
>>>>>>
>>>>>>        avg throughput      135.299 => 138.397  by +2.3%
>>>>>>        avg IOPS            986.969 => 903.344  by -8.5%
>>>>>>
>>>>>> The IOPS is a bit weird.
>>>>>>
>>>>>> Summaries:
>>>>>> - this patch improves RAID throughput by +2.3% on average
>>>>>> - after this patch, 2MB readahead performs slightly better
>>>>>>  (by 1-2%) than 512KB readahead
>>>>> and the most important one:
>>>>> - 64 max_sectors_kb performs much better than 256 max_sectors_kb, by
>>>>> ~30% !
>>>> Yes, I've just wanted to point it out ;)
>>> OK, now I tend to agree on decreasing max_sectors_kb and increasing
>>> read_ahead_kb. But before actually trying to push that idea I'd like
>>> to
>>> - do more benchmarks
>>> - figure out why context readahead didn't help SCST performance
>>>  (previous traces show that context readahead is submitting perfect
>>>   large io requests, so I wonder if it's some io scheduler bug)
>> Because, as we found out, without your http://lkml.org/lkml/2009/5/21/319
>> patch read-ahead was nearly disabled, hence there were no difference which
>> algorithm was used?
>>
>> Ronald, can you run the following tests, please? This time with 2 hosts,
>> initiator (client) and target (server) connected using 1 Gbps iSCSI. It
>> would be the best if on the client vanilla 2.6.29 will be ran, but any other
>> kernel will be fine as well, only specify which. Blockdev-perftest should be
>> ran as before in buffered mode, i.e. with "-a" switch.
> 
> I could, but: only the first 'dd' run of blockdev-perftest will have
> any value, since all others will be served from the target's cache,
> won't that make the results pretty much useless (?). Are you sure this
> is what you want me to test?

Hmm, I forgot about this.. Can you setup possibility to automatically 
ssh from the client to the server and modify drop_caches() function in 
blockdev-perftest on the client so it will instead of

   sync
   echo 3 > /proc/sys/vm/drop_caches

do

   ssh root@target "sync; echo 3 > /proc/sys/vm/drop_caches"

Thanks,
Vlad