Horrible IO performance with SCSI drives

[Argo]

We get this IBM xSeries servers at work that we send to customers and I've noticed a horrible IO performance on those disks. They have LSI 1020/1030 ultra390 adapter with 1 10k 36GB disk and 2 15k 72gb disks. Yet for some reason the perfomance of those disks is really bad. My local IDE drive is about 20 times faster. I tried enabling write caching but that didn't help. Can anybody suggest anything?

 

[beatle]

20k rpm? Who makes that and where'd you get it?

 

[Argo]

Originally posted by: beatle
20k rpm? Who makes that and where'd you get it?

Oops, I meant 10k.

 

[Zepper]

Are you using the correct cable (U320 LVD certified) and termination? Have the drives' proper parameters been punched into the adapter setup? Did someone mistakenly jumper the drives for SE operation? So check all the drive jumpers, cable/terminator and adapter settings. It has to be something among those.
.bh.

:moon: Darn it gets dark early now...

 

[Sideswipe001]

How are you testing them? Is this a benchmark program or just subjective "it feels slow"?

 

[Argo]

Originally posted by: Sideswipe001
How are you testing them? Is this a benchmark program or just subjective "it feels slow"?

Db2 insert rate. Each of the three drives gets less than 100 rows per second, my office PC with 10k SATA drive gets around 1000 inserts per second. Also, raw write speed (i.e. copying 4GB file) gets 4MB/s on the server and 20MB/s on my pc. And oh, yeah, it does feel sluggish.

 

[mechBgon]

Have you asked IBM about it yet?

 

[imported_michaelpatrick33]

Originally posted by: Argo
We get this IBM xSeries servers at work that we send to customers and I've noticed a horrible IO performance on those disks. They have LSI 1020/1030 ultra390 adapter with 1 10k 36GB disk and 2 15k 72gb disks. Yet for some reason the perfomance of those disks is really bad. My local IDE drive is about 20 times faster. I tried enabling write caching but that didn't help. Can anybody suggest anything?

sig. I'm your dinh. May ka guide you to where you've been again with the Wisdom to not have your useless bag of silicon and transistors get squashed by your ignorance. (This is not a post or attack just Darktower fun) His sig has a Darktower reference

I would definitely make sure the hardrives have the correct cable and are you using a 32bit/33mghrtz slot card for them or onboard SCSI etc. That can cause performance penalties by saturation of the bus but that doesn't sound right to me. I would definitely look at the adaptor settings as well and make sure they are not running in SCSI 1 mode or something equally insane.

 

[Argo]

Cool, I'll ask our IT guy to check some of this things out, maybe even call IBM. The funny thing we have 3 of this servers, straight from IBM 2 run linux and 1 runs windows 2000. Windows and one of the linux servers have this problem, as for the 2nd linux server it has oracle so I can't easily compare performance.

 

[sunase]

There's actually a step further you can do beyond setting write caching in Windows called power protected write caching. This prevents Windows from sending write through commands on various occasions, which normally force the disk to commit them as soon as possible. On IDE drives write through isn't supported, but it is on SCSI so this becomes a factor. So setting power protected write caching obviously increases the chance of data loss on a crash/etc, but no more so than using IDE with normal write caching would.

Actually I just realized you may not be using Windows at all, but whatever. ^^

 

[mechBgon]

Also, if you're doing any customization to the IBMs, what other stuff might you be plugging into the PCI-X bus that the controller is riding?

 

[Argo]

Well, we narrowed it down to write performance on the scsi drives under windows 2000. Trying to figure out if we can resolve that or we'll need to upgrade to 2003.

 

[Sideswipe001]

Have you updated your controller's drivers?

 

[Argo]

Originally posted by: Sideswipe001
Have you updated your controller's drivers?

Just did, nothing changed.

This is an article from MS on the problem, seems weird.