Hi Galen,<br><br>I have tried your suggestion and mounted OST directly on /dev/sd<x> devices but that didn't help and I/O is still being fragmented.<br><br>Best regards,<br><br>Wojciech<br><br><div class="gmail_quote">
On 10 June 2011 14:25, Shipman, Galen M. <span dir="ltr"><<a href="mailto:gshipman@ornl.gov">gshipman@ornl.gov</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin: 0pt 0pt 0pt 0.8ex; border-left: 1px solid rgb(204, 204, 204); padding-left: 1ex;">
Wojciech,<br>
<br>
We have seen similar issues with DM-Multipath. Can you experiment with going straight to the block device without DM-Multipath?<br>
<br>
Thanks,<br>
<br>
Galen<br>
<div><div></div><div class="h5"><br>
On Jun 10, 2011, at 8:00 AM, Wojciech Turek wrote:<br>
<br>
> Hi Kevin,<br>
><br>
> Thanks for very helpful answer. I tried your suggestion and recompiled the<br>
> mpt2sas driver with the following changes:<br>
><br>
> --- mpt2sas_base.h 2010-01-16 20:57:30.000000000 +0000<br>
> +++ new_mpt2sas_base.h 2011-06-10 12:53:35.000000000 +0100<br>
> @@ -83,13 +83,13 @@<br>
> #ifdef CONFIG_SCSI_MPT2SAS_MAX_SGE<br>
> #if CONFIG_SCSI_MPT2SAS_MAX_SGE < 16<br>
> #define MPT2SAS_SG_DEPTH 16<br>
> -#elif CONFIG_SCSI_MPT2SAS_MAX_SGE > 128<br>
> -#define MPT2SAS_SG_DEPTH 128<br>
> +#elif CONFIG_SCSI_MPT2SAS_MAX_SGE > 256<br>
> +#define MPT2SAS_SG_DEPTH 256<br>
> #else<br>
> #define MPT2SAS_SG_DEPTH CONFIG_SCSI_MPT2SAS_MAX_SGE<br>
> #endif<br>
> #else<br>
> -#define MPT2SAS_SG_DEPTH 128 /* MAX_HW_SEGMENTS */<br>
> +#define MPT2SAS_SG_DEPTH 256 /* MAX_HW_SEGMENTS */<br>
> #endif<br>
><br>
> #if defined(TARGET_MODE)<br>
><br>
> However I can still that almost 50% of writes and slightly over 50% of reads<br>
> falls under 512K I/Os<br>
> I am using device-mapper-multipath to manage active/passive paths do you<br>
> think that could have something to do with the I/O fragmentation?<br>
><br>
> Best regards,<br>
><br>
> Wojciech<br>
><br>
> On 8 June 2011 17:30, Kevin Van Maren <<a href="mailto:kevin.van.maren@oracle.com">kevin.van.maren@oracle.com</a>> wrote:<br>
><br>
>> Yep, with 1.8.5 the problem is most likely in the (mpt2sas) driver, not in<br>
>> the rest of the kernel. Driver limits are not normally noticed by<br>
>> (non-Lustre) people, because the default kernel limits IO to 512KB.<br>
>><br>
>> May want to see Bug 22850 for the changes required eg, for the Emulex/lpfc<br>
>> driver.<br>
>><br>
>> Glancing at the stock RHEL5 kernel, it looks like the issue is<br>
>> MPT2SAS_SG_DEPTH, which is limited to 128. This appears to be set to match<br>
>> the default kernel limit, but it is possible there is also a driver/HW<br>
>> limit. You should be able to increase that to 256 and see if it works...<br>
>><br>
>><br>
>> Also note that the size buckets are power-of-2, so a "1MB" entry is any IO<br>
>>> 512KB and <= 1MB.<br>
>><br>
>> If you can't get the driver to reliably do full 1MB IOs, change to a 64KB<br>
>> chunk and set max_sectors_kb to 512. This will help ensure you get aligned,<br>
>> full-stripe writes.<br>
>><br>
>> Kevin<br>
>><br>
>><br>
>><br>
>> Wojciech Turek wrote:<br>
>><br>
>>> I am setting up a new lustre filesystem using LSI engenio based disk<br>
>>> enclosures with integrated dual RAID controllers. I configured disks<br>
>>> into 8+2 RAID6 groups using 128kb segment size (chunk size). This<br>
>>> hardware uses mpt2sas kernel module on the Linux host side. I use the<br>
>>> whole block device for an OST (to avoid any alignment issues). When<br>
>>> running sgpdd-survey I can see high through numbers (~3GB/s write,<br>
>>> 5GB/s read), Also controllers stats show that number of IOPS = number<br>
>>> of MB/s. However as soon as I put ldiskfs on the OSTs, obdfilter shows<br>
>>> slower results (~2GB/s write , ~2GB/s read ) and controller stats show<br>
>>> more then double IOPS than MB/s. Looking at output from iostat -m -x 1<br>
>>> and brw_stats I can see that a large number of I/O operations are<br>
>>> smaller than 1MB, mostly 512kb. I know that there was some work done<br>
>>> on optimising the kernel block device layer to process 1MB I/O<br>
>>> requests and that those changes were committed to Lustre 1.8.5. Thus I<br>
>>> guess this I/O chopping happens below the Lustre stack, maybe in the<br>
>>> mpt2sas driver?<br>
>>><br>
>>> I am hoping that someone in Lustre community can shed some light on to<br>
>>> my problem.<br>
>>><br>
>>> In my setup I use:<br>
>>> Lustre 1.8.5<br>
>>> CentOS-5.5<br>
>>><br>
>>> Some parameters I tuned from defaults in CentOS:<br>
>>> deadline I/O scheduler<br>
>>><br>
>>> max_hw_sectors_kb=4096<br>
>>> max_sectors_kb=1024<br>
>>><br>
>>><br>
>>> brw_stats output<br>
>>> --<br>
>>><br>
>>> find /proc/fs/lustre/obdfilter/ -name "testfs-OST*" | while read ost;<br>
>>> do cat $ost/brw_stats ; done | grep "disk I/O size" -A9<br>
>>><br>
>>> disk I/O size ios % cum % | ios % cum %<br>
>>> 4K: 206 0 0 | 521 0 0<br>
>>> 8K: 224 0 0 | 595 0 1<br>
>>> 16K: 105 0 1 | 479 0 1<br>
>>> 32K: 140 0 1 | 1108 1 3<br>
>>> 64K: 231 0 1 | 1470 1 4<br>
>>> 128K: 536 1 2 | 2259 2 7<br>
>>> 256K: 1762 3 6 | 5644 6 14<br>
>>> 512K: 31574 64 71 | 30431 35 50<br>
>>> 1M: 14200 28 100 | 42143 49 100<br>
>>> --<br>
>>> disk I/O size ios % cum % | ios % cum %<br>
>>> 4K: 187 0 0 | 457 0 0<br>
>>> 8K: 244 0 0 | 598 0 1<br>
>>> 16K: 109 0 1 | 481 0 1<br>
>>> 32K: 129 0 1 | 1100 1 3<br>
>>> 64K: 222 0 1 | 1408 1 4<br>
>>> 128K: 514 1 2 | 2291 2 7<br>
>>> 256K: 1718 3 6 | 5652 6 14<br>
>>> 512K: 32222 65 72 | 29810 35 49<br>
>>> 1M: 13654 27 100 | 42202 50 100<br>
>>> --<br>
>>> disk I/O size ios % cum % | ios % cum %<br>
>>> 4K: 196 0 0 | 551 0 0<br>
>>> 8K: 206 0 0 | 551 0 1<br>
>>> 16K: 79 0 0 | 513 0 1<br>
>>> 32K: 136 0 1 | 1048 1 3<br>
>>> 64K: 232 0 1 | 1278 1 4<br>
>>> 128K: 540 1 2 | 2172 2 7<br>
>>> 256K: 1681 3 6 | 5679 6 13<br>
>>> 512K: 31842 64 71 | 31705 37 51<br>
>>> 1M: 14077 28 100 | 41789 48 100<br>
>>> --<br>
>>> disk I/O size ios % cum % | ios % cum %<br>
>>> 4K: 190 0 0 | 486 0 0<br>
>>> 8K: 200 0 0 | 547 0 1<br>
>>> 16K: 93 0 0 | 448 0 1<br>
>>> 32K: 141 0 1 | 1029 1 3<br>
>>> 64K: 240 0 1 | 1283 1 4<br>
>>> 128K: 558 1 2 | 2125 2 7<br>
>>> 256K: 1716 3 6 | 5400 6 13<br>
>>> 512K: 31476 64 70 | 29029 35 48<br>
>>> 1M: 14366 29 100 | 42454 51 100<br>
>>> --<br>
>>> disk I/O size ios % cum % | ios % cum %<br>
>>> 4K: 209 0 0 | 511 0 0<br>
>>> 8K: 195 0 0 | 621 0 1<br>
>>> 16K: 79 0 0 | 558 0 1<br>
>>> 32K: 134 0 1 | 1135 1 3<br>
>>> 64K: 245 0 1 | 1390 1 4<br>
>>> 128K: 509 1 2 | 2219 2 7<br>
>>> 256K: 1715 3 6 | 5687 6 14<br>
>>> 512K: 31784 64 71 | 31172 36 50<br>
>>> 1M: 14112 28 100 | 41719 49 100<br>
>>> --<br>
>>> disk I/O size ios % cum % | ios % cum %<br>
>>> 4K: 201 0 0 | 500 0 0<br>
>>> 8K: 241 0 0 | 604 0 1<br>
>>> 16K: 82 0 1 | 584 0 1<br>
>>> 32K: 130 0 1 | 1092 1 3<br>
>>> 64K: 230 0 1 | 1331 1 4<br>
>>> 128K: 547 1 2 | 2253 2 7<br>
>>> 256K: 1695 3 6 | 5634 6 14<br>
>>> 512K: 31501 64 70 | 31836 37 51<br>
>>> 1M: 14343 29 100 | 41517 48 100<br>
>>><br>
>>><br>
>>><br>
>><br>
>><br>
</div></div><div class="im">> _______________________________________________<br>
> Lustre-discuss mailing list<br>
> <a href="mailto:Lustre-discuss@lists.lustre.org">Lustre-discuss@lists.lustre.org</a><br>
</div>> blockedhttp://<a href="http://lists.lustre.org/mailman/listinfo/lustre-discuss" target="_blank">lists.lustre.org/mailman/listinfo/lustre-discuss</a><br>
<br>
</blockquote></div><br><br>