[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <4A5238EC.1070505@vlnb.net>
Date: Mon, 06 Jul 2009 21:48:28 +0400
From: Vladislav Bolkhovitin <vst@...b.net>
To: Ronald Moesbergen <intercommit@...il.com>
CC: Wu Fengguang <fengguang.wu@...el.com>,
linux-kernel@...r.kernel.org, akpm@...ux-foundation.org,
kosaki.motohiro@...fujitsu.com, Alan.Brunelle@...com,
hifumi.hisashi@....ntt.co.jp, linux-fsdevel@...r.kernel.org,
jens.axboe@...cle.com, randy.dunlap@...cle.com,
Bart Van Assche <bart.vanassche@...il.com>
Subject: Re: [RESEND] [PATCH] readahead:add blk_run_backing_dev
Ronald Moesbergen, on 07/06/2009 06:37 PM wrote:
> 2009/7/6 Vladislav Bolkhovitin <vst@...b.net>:
>> (Restored the original list of recipients in this thread as I was asked.)
>>
>> Hi Ronald,
>>
>> Ronald Moesbergen, on 07/04/2009 07:19 PM wrote:
>>> 2009/7/3 Vladislav Bolkhovitin <vst@...b.net>:
>>>> Ronald Moesbergen, on 07/03/2009 01:14 PM wrote:
>>>>>>> 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.
>>>>>>
>>>>>> 1. All defaults on the client, on the server vanilla 2.6.29 with
>>>>>> Fengguang's
>>>>>> http://lkml.org/lkml/2009/5/21/319 patch with all default settings.
>>>>>>
>>>>>> 2. All defaults on the client, on the server vanilla 2.6.29 with
>>>>>> Fengguang's
>>>>>> http://lkml.org/lkml/2009/5/21/319 patch with default RA size and 64KB
>>>>>> max_sectors_kb.
>>>>>>
>>>>>> 3. All defaults on the client, on the server vanilla 2.6.29 with
>>>>>> Fengguang's
>>>>>> http://lkml.org/lkml/2009/5/21/319 patch with 2MB RA size and default
>>>>>> max_sectors_kb.
>>>>>>
>>>>>> 4. All defaults on the client, on the server vanilla 2.6.29 with
>>>>>> Fengguang's
>>>>>> http://lkml.org/lkml/2009/5/21/319 patch with 2MB RA size and 64KB
>>>>>> max_sectors_kb.
>>>>>>
>>>>>> 5. All defaults on the client, on the server vanilla 2.6.29 with
>>>>>> Fengguang's
>>>>>> http://lkml.org/lkml/2009/5/21/319 patch and with context RA patch. RA
>>>>>> size
>>>>>> and max_sectors_kb are default. For your convenience I committed the
>>>>>> backported context RA patches into the SCST SVN repository.
>>>>>>
>>>>>> 6. All defaults on the client, on the server vanilla 2.6.29 with
>>>>>> Fengguang's
>>>>>> http://lkml.org/lkml/2009/5/21/319 and context RA patches with default
>>>>>> RA
>>>>>> size and 64KB max_sectors_kb.
>>>>>>
>>>>>> 7. All defaults on the client, on the server vanilla 2.6.29 with
>>>>>> Fengguang's
>>>>>> http://lkml.org/lkml/2009/5/21/319 and context RA patches with 2MB RA
>>>>>> size
>>>>>> and default max_sectors_kb.
>>>>>>
>>>>>> 8. All defaults on the client, on the server vanilla 2.6.29 with
>>>>>> Fengguang's
>>>>>> http://lkml.org/lkml/2009/5/21/319 and context RA patches with 2MB RA
>>>>>> size
>>>>>> and 64KB max_sectors_kb.
>>>>>>
>>>>>> 9. On the client default RA size and 64KB max_sectors_kb. On the server
>>>>>> vanilla 2.6.29 with Fengguang's http://lkml.org/lkml/2009/5/21/319 and
>>>>>> context RA patches with 2MB RA size and 64KB max_sectors_kb.
>>>>>>
>>>>>> 10. On the client 2MB RA size and default max_sectors_kb. On the server
>>>>>> vanilla 2.6.29 with Fengguang's http://lkml.org/lkml/2009/5/21/319 and
>>>>>> context RA patches with 2MB RA size and 64KB max_sectors_kb.
>>>>>>
>>>>>> 11. On the client 2MB RA size and 64KB max_sectors_kb. On the server
>>>>>> vanilla
>>>>>> 2.6.29 with Fengguang's http://lkml.org/lkml/2009/5/21/319 and context
>>>>>> RA
>>>>>> patches with 2MB RA size and 64KB max_sectors_kb.
>>>>> Ok, done. Performance is pretty bad overall :(
>>>>>
>>>>> The kernels I used:
>>>>> client kernel: 2.6.26-15lenny3 (debian)
>>>>> server kernel: 2.6.29.5 with blk_dev_run patch
>>>>>
>>>>> And I adjusted the blockdev-perftest script to drop caches on both the
>>>>> server (via ssh) and the client.
>>>>>
>>>>> The results:
>>>>>
>>> ... previous results ...
>>>
>>>> Those are on the server without io_context-2.6.29 and readahead-2.6.29
>>>> patches applied and with CFQ scheduler, correct?
>>>>
>>>> Then we see how reorder of requests caused by many I/O threads submitting
>>>> I/O in separate I/O contexts badly affect performance and no RA,
>>>> especially
>>>> with default 128KB RA size, can solve it. Less max_sectors_kb on the
>>>> client
>>>> => more requests it sends at once => more reorder on the server => worse
>>>> throughput. Although, Fengguang, in theory, context RA with 2MB RA size
>>>> should considerably help it, no?
>>>>
>>>> Ronald, can you perform those tests again with both io_context-2.6.29 and
>>>> readahead-2.6.29 patches applied on the server, please?
>>> Hi Vlad,
>>>
>>> I have retested with the patches you requested (and got access to the
>>> systems today :) ) The results are better, but still not great.
>>>
>>> client kernel: 2.6.26-15lenny3 (debian)
>>> server kernel: 2.6.29.5 with io_context and readahead patch
>>>
>>> 5) client: default, server: default
>>> blocksize R R R R(avg, R(std R
>>> (bytes) (s) (s) (s) MB/s) ,MB/s) (IOPS)
>>> 67108864 18.303 19.867 18.481 54.299 1.961 0.848
>>> 33554432 18.321 17.681 18.708 56.181 1.314 1.756
>>> 16777216 17.816 17.406 19.257 56.494 2.410 3.531
>>> 8388608 18.077 17.727 19.338 55.789 2.056 6.974
>>> 4194304 17.918 16.601 18.287 58.276 2.454 14.569
>>> 2097152 17.426 17.334 17.610 58.661 0.384 29.331
>>> 1048576 19.358 18.764 17.253 55.607 2.734 55.607
>>> 524288 17.951 18.163 17.440 57.379 0.983 114.757
>>> 262144 18.196 17.724 17.520 57.499 0.907 229.995
>>> 131072 18.342 18.259 17.551 56.751 1.131 454.010
>>> 65536 17.733 18.572 17.134 57.548 1.893 920.766
>>> 32768 19.081 19.321 17.364 55.213 2.673 1766.818
>>> 16384 17.181 18.729 17.731 57.343 2.033 3669.932
>>>
>>> 6) client: default, server: 64 max_sectors_kb, RA default
>>> blocksize R R R R(avg, R(std R
>>> (bytes) (s) (s) (s) MB/s) ,MB/s) (IOPS)
>>> 67108864 21.790 20.062 19.534 50.153 2.304 0.784
>>> 33554432 20.212 19.744 19.564 51.623 0.706 1.613
>>> 16777216 20.404 19.329 19.738 51.680 1.148 3.230
>>> 8388608 20.170 20.772 19.509 50.852 1.304 6.356
>>> 4194304 19.334 18.742 18.522 54.296 0.978 13.574
>>> 2097152 19.413 18.858 18.884 53.758 0.715 26.879
>>> 1048576 20.472 18.755 18.476 53.347 2.377 53.347
>>> 524288 19.120 20.104 18.404 53.378 1.925 106.756
>>> 262144 20.337 19.213 18.636 52.866 1.901 211.464
>>> 131072 19.199 18.312 19.970 53.510 1.900 428.083
>>> 65536 19.855 20.114 19.592 51.584 0.555 825.342
>>> 32768 20.586 18.724 20.340 51.592 2.204 1650.941
>>> 16384 21.119 19.834 19.594 50.792 1.651 3250.669
>>>
>>> 7) client: default, server: default max_sectors_kb, RA 2MB
>>> blocksize R R R R(avg, R(std R
>>> (bytes) (s) (s) (s) MB/s) ,MB/s) (IOPS)
>>> 67108864 17.767 16.489 16.949 60.050 1.842 0.938
>>> 33554432 16.777 17.034 17.102 60.341 0.500 1.886
>>> 16777216 18.509 16.784 16.971 58.891 2.537 3.681
>>> 8388608 18.058 17.949 17.599 57.313 0.632 7.164
>>> 4194304 18.286 17.648 17.026 58.055 1.692 14.514
>>> 2097152 17.387 18.451 17.875 57.226 1.388 28.613
>>> 1048576 18.270 17.698 17.570 57.397 0.969 57.397
>>> 524288 16.708 17.900 17.233 59.306 1.668 118.611
>>> 262144 18.041 17.381 18.035 57.484 1.011 229.934
>>> 131072 17.994 17.777 18.146 56.981 0.481 455.844
>>> 65536 17.097 18.597 17.737 57.563 1.975 921.011
>>> 32768 17.167 17.035 19.693 57.254 3.721 1832.127
>>> 16384 17.144 16.664 17.623 59.762 1.367 3824.774
>>>
>>> 8) client: default, server: 64 max_sectors_kb, RA 2MB
>>> blocksize R R R R(avg, R(std R
>>> (bytes) (s) (s) (s) MB/s) ,MB/s) (IOPS)
>>> 67108864 20.003 21.133 19.308 50.894 1.881 0.795
>>> 33554432 19.448 20.015 18.908 52.657 1.222 1.646
>>> 16777216 19.964 19.350 19.106 52.603 0.967 3.288
>>> 8388608 18.961 19.213 19.318 53.437 0.419 6.680
>>> 4194304 18.135 19.508 19.361 53.948 1.788 13.487
>>> 2097152 18.753 19.471 18.367 54.315 1.306 27.158
>>> 1048576 19.189 18.586 18.867 54.244 0.707 54.244
>>> 524288 18.985 19.199 18.840 53.874 0.417 107.749
>>> 262144 19.064 21.143 19.674 51.398 2.204 205.592
>>> 131072 18.691 18.664 19.116 54.406 0.594 435.245
>>> 65536 18.468 20.673 18.554 53.389 2.729 854.229
>>> 32768 20.401 21.156 19.552 50.323 1.623 1610.331
>>> 16384 19.532 20.028 20.466 51.196 0.977 3276.567
>>>
>>> 9) client: 64 max_sectors_kb, default RA. server: 64 max_sectors_kb, RA
>>> 2MB
>>> blocksize R R R R(avg, R(std R
>>> (bytes) (s) (s) (s) MB/s) ,MB/s) (IOPS)
>>> 67108864 16.458 16.649 17.346 60.919 1.364 0.952
>>> 33554432 16.479 16.744 17.069 61.096 0.878 1.909
>>> 16777216 17.128 16.585 17.112 60.456 0.910 3.778
>>> 8388608 17.322 16.780 16.885 60.262 0.824 7.533
>>> 4194304 17.530 16.725 16.756 60.250 1.299 15.063
>>> 2097152 16.580 17.875 16.619 60.221 2.076 30.110
>>> 1048576 17.550 17.406 17.075 59.049 0.681 59.049
>>> 524288 16.492 18.211 16.832 59.718 2.519 119.436
>>> 262144 17.241 17.115 17.365 59.397 0.352 237.588
>>> 131072 17.430 16.902 17.511 59.271 0.936 474.167
>>> 65536 16.726 16.894 17.246 60.404 0.768 966.461
>>> 32768 16.662 17.517 17.052 59.989 1.224 1919.658
>>> 16384 17.429 16.793 16.753 60.285 1.085 3858.268
>>>
>>> 10) client: default max_sectors_kb, 2MB RA. server: 64 max_sectors_kb, RA
>>> 2MB
>>> blocksize R R R R(avg, R(std R
>>> (bytes) (s) (s) (s) MB/s) ,MB/s) (IOPS)
>>> 67108864 17.601 18.334 17.379 57.650 1.307 0.901
>>> 33554432 18.281 18.128 17.169 57.381 1.610 1.793
>>> 16777216 17.660 17.875 17.356 58.091 0.703 3.631
>>> 8388608 17.724 17.810 18.383 56.992 0.918 7.124
>>> 4194304 17.475 17.770 19.003 56.704 2.031 14.176
>>> 2097152 17.287 17.674 18.492 57.516 1.604 28.758
>>> 1048576 17.972 17.460 18.777 56.721 1.689 56.721
>>> 524288 18.680 18.952 19.445 53.837 0.890 107.673
>>> 262144 18.070 18.337 18.639 55.817 0.707 223.270
>>> 131072 16.990 16.651 16.862 60.832 0.507 486.657
>>> 65536 17.707 16.972 17.520 58.870 1.066 941.924
>>> 32768 17.767 17.208 17.205 58.887 0.885 1884.399
>>> 16384 18.258 17.252 18.035 57.407 1.407 3674.059
>>>
>>> 11) client: 64 max_sectors_kb, 2MB. RA server: 64 max_sectors_kb, RA 2MB
>>> blocksize R R R R(avg, R(std R
>>> (bytes) (s) (s) (s) MB/s) ,MB/s) (IOPS)
>>> 67108864 17.993 18.307 18.718 55.850 0.902 0.873
>>> 33554432 19.554 18.485 17.902 54.988 1.993 1.718
>>> 16777216 18.829 18.236 18.748 55.052 0.785 3.441
>>> 8388608 21.152 19.065 18.738 52.257 2.745 6.532
>>> 4194304 19.131 19.703 17.850 54.288 2.268 13.572
>>> 2097152 19.093 19.152 19.509 53.196 0.504 26.598
>>> 1048576 19.371 18.775 18.804 53.953 0.772 53.953
>>> 524288 20.003 17.911 18.602 54.470 2.476 108.940
>>> 262144 19.182 19.460 18.476 53.809 1.183 215.236
>>> 131072 19.403 19.192 18.907 53.429 0.567 427.435
>>> 65536 19.502 19.656 18.599 53.219 1.309 851.509
>>> 32768 18.746 18.747 18.250 55.119 0.701 1763.817
>>> 16384 20.977 19.437 18.840 51.951 2.319 3324.862
>> The results look inconsistently with what you had previously (89.7 MB/s).
>> How can you explain it?
>
> I had more patches applied with that test: (scst_exec_req_fifo-2.6.29,
> put_page_callback-2.6.29) and I used a different dd command:
>
> dd if=/dev/sdc of=/dev/zero bs=512K count=2000
>
> But all that said, I can't reproduce speeds that high now. Must have
> made a mistake back then (maybe I forgot to clear the pagecache).
If you forgot to clear the cache, you would had had the wire throughput
(110 MB/s) or more.
>> I think, most likely, there was some confusion between the tested and
>> patched versions of the kernel or you forgot to apply the io_context patch.
>> Please recheck.
>
> The tests above were definitely done right, I just rechecked the
> patches, and I do see an average increase of about 10MB/s over an
> unpatched kernel. But overall the performance is still pretty bad.
Have you rebuild and reinstall SCST after patching kernel?
> Ronald.
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@...r.kernel.org
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/
>
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
Powered by blists - more mailing lists