LVM Striping size Question

[Gordon_3]

Hi all,

Currently I hit a problem, the main problem is one of the Sybase server, which block size is 2K by default, generate huge I/O ( read intensive ) all the day. the LV is 5GB, and no striping, it can grow up to 2XXX I/O per sec all day. Now I decide to use LVM to stripe across more spindle to help to ease the situation. Have read some previous post that LVM stripe size 64K should be a best choice as LVM for HP, as LVM "group" I/O in 64K chunk for disk I/O. I wonder is it mean that after setting this LVM striping, number of I/O can be reduce ?? Pls enlighten and thx.

Gordon

[Stefan Farrelly]

Certainly striping over all your spindles should help a lot. As for stripe size, remember vxfs filesystems use a 1k blocksize.

Seeing as sybase is using 2k I would try a stripsize of 8k first, test/benchmark it, and then move up to 16,32,64. Usually a 64k stripe size is best but for specific applications, like Sybase, you really need to test it first to be sure which is optimal.

Read this, it explains everything. Even though its for Oracle the same applies for most databases;

http://www.dutchworks.nl/htbin/hpsysadmin?h=3&dn=88193&q=vxfs%20blocksi%7ae&fh

[Leif Halvarsson_2]

Hi,
Stiping a volume group will perhaps not reduce the numbers of IOPS but a striped volume group can handle more IOPS. It seems strange with IOPS at 2000/s if using a ordinary disk. A 10000 rpm disk will, at best give about 250 IOPS. Perhaps most of your IO is file chache.

The best way to get very high IO performance is to use a hardware RAID controller.

[Ian Dennison_1]

Gordon,

Oh boy, why not ask something really controversial, like who shot JFK!

Everyone has their own opinion on how to set up striping and how to determine stripe size. Estimates range from 8K (apparently best for Oracle) to 1MB (the S.A.M.E. method).

Attached are 2 references to forum postings; the first gives you an idea about the various discussions, the second (look for the posting by Tim Fulford) is the one I used to determine my optimum stripe size for my Oracle Database.

http://forums.itrc.hp.com/cm/QuestionAnswer/1,,0x86727d4cf554d611abdb0090277a778c,00.html

http://forums.itrc.hp.com/cm/QuestionAnswer/1,,0x998c402f24d5d61190050090279cd0f9,00.html

Share and Enjoy! Ian

[Dietmar Konermann]

Something completely different... is your Sybase database not configured to use the asyncdsk driver (/dev/async)? Doing such a huge number of 2K IOs is a really stupid thing, at least from a performance perspective.

Best regards...
Dietmar.

[Gordon_3]

HI all professionals,

That's why I like to use Forum as here got so many professional viewpoints ^_^ thx again, yeah first we are using XP1024 and so that's why can get such "monster I/O" rate, our config is 8 X 15000 rpm hardware striping for 1 LV. And our Sybase is using async I/O now, I will review the links u guys post before raising more Q, anyway thx all u guys. It really "open" my eyes..

Gordon

[Gordon_3]

Ar one more point I miss, checking the Perfview and Performance anaylser, found high percentage is Sequential read I/O ( is it due to our Sybase is Raw device ?? ), only very little is write I/O, that's why I wanna try 64K at first as suppose it's benefit for sequential write, right?

Gordon

[Ian Dennison_1]

Gordon,

I may be jumping the gun, but I suppose with an XP Array, additional considerations will be the backend layout (across various ACPs) and the frontend capability of the CHIP and associated FC switches.

Do you have an rough indication of how busy your CHIP pairs are? (performance agent does not really provide historical information on it)?

Share and Enjoy! Ian

[Gordon_3]

HI Ian,

U mean the Performance Analyser figure about the CHIP and ACP is not correct? Well coz both utilization is very low ( at most 15% util ) and not a bottlenect point, thx.

Gordon

[Ian Dennison_1]

Gordon,

Actually I meant that the Charts option on Performance Advisor does not allow me to see CHIP Utilisation. I have to drill into the Components section to see the information, and combine it to see all CHIP Pairs side-by-side (Our CHIPs apparently get really loaded during the night for backups, but it takes half a day to generate the graphs individually).

Share and Enjoy! Ian