9550sx-8lp read performance problems....

/dev/null

/dev/null

[H]F Junkie
Joined
Mar 31, 2001
Messages
15,181
I have 5 * 400GB seagate 7200.8 drives in raid5.

I have the latest bios & drivesr for the card

My sustained read performance on large files is ~ 10-12MB/second

I can write at 50-60MB/second...uhhhhh ?

Any ideas?
 
In a word: no. No ideas. Did you test a single drive? What app is giving you those numbers? What's HDTach show?

 
Well I'm using a windows based md5 sum program. Using this on single drives yields 40-50MB/second. In a software raid5 read > 100MB/second. The 3ware SHOULD be doing much better.

This is ona GA-2CEWH gigabyte board in a pci-x slot.

Rob
 
Try HDTach; it gets more repeatable results in my experience. The MD5 program could be limited by bad coding or CPU usage.

 
Well, the md5 program will read at > 200MB/second (as measured by perfmon) when I had this in windows software raid5, so I don't think it's the MD5 program.

Could my bad performance be because I have left the raid5 as a "basic" disk in windows?

I swear I read something about basic disks having crappy performance compared to dynamic disks.

Anyone?

I will post hdtach numbers tonight. Which test should I run?
 
Just the short test will suffice. I haven't heard anything about basic versus dynamic disks, performance-wise. What I have heard is it's harder to recover from crashes with them, because third-party software people don't understand the format as reliably as plain dos partitions plus NTFS.

 
That looks a bit low, but at least it's showing 50 MB/s. What chunk size did you use, and are you using standard NTFS? I really don't know why it's so bad; those are good disks for high STR, and there's nothing else I can think of that should be slowing you down.

 
unhappy_mage said:
That looks a bit low, but at least it's showing 50 MB/s. What chunk size did you use, and are you using standard NTFS? I really don't know why it's so bad; those are good disks for high STR, and there's nothing else I can think of that should be slowing you down.

Click to expand...

Yep I'm using standard NTFS. I think my problem is probably cluster vs stripe size. I have a 4kb cluster size and 256kb stripe size. Would this lead to bad performance?

3ware also had me run the "atto" benchmark.

Click for Results
 
No, that shouldn't be a problem. Holy crap those write figures look bad, though. 50 MB/s? I can do that with 3 300s and software raid 5 under Linux. Looks like 3ware gets to stay on my bad list.

Is the card in write-through or write-back mode? (see definition) I wouldn't imagine it'd make that much difference, but I guess it could...

 
unhappy_mage said:
No, that shouldn't be a problem. Holy crap those write figures look bad, though. 50 MB/s? I can do that with 3 300s and software raid 5 under Linux. Looks like 3ware gets to stay on my bad list.

Is the card in write-through or write-back mode? (see definition) I wouldn't imagine it'd make that much difference, but I guess it could...

Click to expand...

From both 3wares & storagereviews site, I was using a stripe size that was better for random access than sequential. I have changed this.

I am now trying to figure out the proper file system "cluster" size to format the drive as.

Suggestions ? :)
 
The default of 4 or 8k (whichever; I don't remember) shouldn't pose a problem. The filesystem shouldn't be the issue, and HDTach ignores it - it reads blocks directly from the disk, not files from the filesystem. Use that as your meter of whether you've fixed it or not.

 
Well, that's better... I guess... I mean, the average read is a lot better, but the huge zigzags still leave me wondering. Is this the short test or the long one? Even if the longer test would smooth the graph out a little, I think it'd be a red herring - the performance still has the spiky performance, so getting a smoother graph would be misleading. If it's still like that after rebuild finishes, I'd look into some more possible solutions. I don't know any other than "play with chunk size some more" but that'd make a good starting point, I guess.

 
unhappy_mage said:
Well, that's better... I guess... I mean, the average read is a lot better, but the huge zigzags still leave me wondering. Is this the short test or the long one? Even if the longer test would smooth the graph out a little, I think it'd be a red herring - the performance still has the spiky performance, so getting a smoother graph would be misleading. If it's still like that after rebuild finishes, I'd look into some more possible solutions. I don't know any other than "play with chunk size some more" but that'd make a good starting point, I guess.

Click to expand...

What I find interesting is the IO performance.

I downloaded storage reviews iometer file from their 2002 tests (not sure if anything is newer). My exact drive, according to their review got something like 128 IO's on a 1 IO outsanding test.

My rig pulled just over 700 with 5 drives :eek: The top performing drive for that benchmark did 372 IO's w/1 outstanding. It's a 15krpm 147G u320 scsi drive that sells for $800 at zzf
 
You must log in or register to reply here.
Back
Top