up vote 1 down vote favorite
share [g+] share [fb]

I'm looking for a recommendation on block size and RAID0 stripe size.

We have a RAID0 array consisting of six SAS drives. The array is a workspace for SAS programs (I know, the various SAS references are confusing). We have a handful of these workspaces set up, and so when a user submits a SAS program on our server, their program gets it's own dedicated workspace.

Our current set up:

OS: SLES 11

Filesystem: ext3

Filesystem block size: 4k

RAID0 stripe size: 64k

There are also application level settings for SAS for block size (BLKSIZE), currently set to 256 bytes -- this must have been a default value.

Is anyone familar with setting up workspaces for SAS processing that could provide feedback on our current settings? We are looking to increase performance and speed up program run times. Let me know if I've left out any information.

link|improve this question
100% accept rate
feedback

4 Answers

up vote 1 down vote accepted

Not exactly what you were asking for, but waisting performance by using ext3 is really bad. You should really look into using ext4 or XFS.

No experience with SAS, so it's impossible for me to answer how large stripe or block size you should use. It all depends on how big the files are.

link|improve this answer
Good point on the filesystems. I've been benchmarking various filesystems these past few days, and was happy to find a 10% performance increase, on average, when using xfs instead of ext3. I'd love to check out ext4, but I'm not sure it's ready for production use. – Banjer Jun 30 '10 at 13:02
1  
ext4 is ready for production use in Ubuntu 10.4 and the latest RedHat/Fedora releases AFAIK. But yeah - every time i've benchamrked our applications and databases on Linux the winner is XFS, so I've just defaulted to it from now on. – pauska Jun 30 '10 at 14:04
feedback
up vote 1 down vote

Without knowing the read-write profile of the application it's pretty impossible to recommend an improved setup. RAID10 is faster if it's 90/10 (R/W), and RAID 0 is faster if it's 10/90.

link|improve this answer
feedback
up vote 1 down vote

Regardless on the app settings, your hardware setup is a serious issue.

The failure probability on RAID 0 is worse than on a single disk. Assuming a 5% failure probability within X years, putting 6 disks into the equation gives you a probability of (1-(1-0.05)^6) ~= 26.5%.

Sacrifice some of your capacity and go with a RAID 10 instead. Otherwise, expect your app to fall over fast.

Ok, assuming you're using these as "scratch space": if you're using software RAID, would it be possible to switch to hardware RAID controllers?

link|improve this answer
Considering this is a processing computer, they probably keep a backup of the system, a backup of the data going in, and just assume it'll fail once every 4 years on average. If the system is built for fast recovery (or full server redundancy) there's no reason to be that concerned. – Chris S♦ Jun 29 '10 at 23:47
@chris OK, the TLAs were numbing my brain; if they're being used as "scratch space" then performance is possibly more important than reliability. Answer changed accordingly. – Andrew Jun 30 '10 at 1:10
Good point, a good hardware RAID card will almost always beat software RAID. – Chris S♦ Jun 30 '10 at 2:28
Thanks for the responses. Right, it is just scratch space, i.e. just a workspace used for IO during program execution. Our setup is all about performance, no backups or redundancy needed whatsoever. We use 3ware HW RAID controllers. – Banjer Jun 30 '10 at 13:00
feedback
up vote 0 down vote

It all depends on how you're using it. Drives have very good sustained read/write speed, and relatively terrible seek speed. You want a drive to be able to read or write as much in a single stretch as possible. However, if you're only working on one file at a time, then you want to be able to split the work over as many drives as possible. If your stripe size is too large, then you won't be able to use as many drives.

You'll probably have to try a few setups to find your sweet spot. Don't guess; benchmark. Test it, change it, test it again, and compare. Repeat until satisfied.

link|improve this answer
feedback

Your Answer

 
or
required, but never shown
discard

Not the answer you're looking for? Browse other questions tagged or ask your own question.