Page 1 of 1

TR's IDE RAID round-up

Posted: Thu Dec 05, 2002 2:37 pm
by Mr Bill
TR's IDE RAID round-up by Geoff Gasior December 4, 2002
http://www.tech-report.com/reviews/2002 ... dex.x?pg=1
This is a very interesting review. I wonder Geoff if you would consider adding a comparison of software raid to what you have already tested? The reason I ask this is because NT and W2Kpro let you set up software raid logical drives which do not require dedication of the entire drive array to a single raid type. Among other things this means that drive failure destroys only the software raid logical drive and other partitions on that particular drive but not the partitions on the rest of the drives. I realize that software raid is CPU intensive but I think it would be a worthwhile addition to your article.

Would you mind summarizing the IOMeter settings you used in your tests? I am not quite sure how you set them up and I would like to try my software raid 0 drive against the same tests.

Posted: Thu Dec 05, 2002 3:23 pm
by Dissonance
I considered doing software RAID, but my choices for RAID levels was pretty limited by the operating system used. Introducing a new operating system just to do software RAID testing might have otherwise skewed the results because of the different OS versions, though when I get some time it's something I'll probably look at in another article.

As for IOMeter, which specific details did you want? The test pattern specifics are bundled with IOMeter and can also be found on StorageReview. As for the test setup, I used a 30-second warmup time for each test, and ran each test for two minutes. I used the exponential stepping setting to test load scaling.

Posted: Thu Dec 05, 2002 4:09 pm
by Mr Bill
Well, I wondering what specific settings you used to do the tests. As I sit here looking at the IOMeter manager I see a default of two workers probably because I have a duallie setup.

First tab "Disk Targets" what did you enter for "# of Outstanding I/O's"?
Did you enter anything for "Test Connection Rate"?

I assume Network Targets not used

What Access specifications did you load?

Tab "Test Setup" what did you ramp to create your load? Was that "Workers", "Targets" or "# of Outstanding I/O's"?

Does the number of workers in the "Topology" matter in the context of the "Test Setup" Tab?

Thanks for answering

Posted: Fri Dec 06, 2002 3:44 am
by Dissonance
I can pull together some screen shots and email them to you if you'd like, Mr. Bill. Otherwise I can try to rig up my server to host them and just post links here. I'm assuming that a screen shot of each configuration tab for IOMeter should suffice?

Also, you might not want to run the tests on a drive that you're currently using, since running the test involves 'preparing' the drives, which pretty much means filling them to capacity.

Posted: Fri Dec 06, 2002 6:56 am
by NeXus 6
Hey Diss -

Any chance on doing a future RAID article, showing the speed differences between using different stripe and cluster sizes? Personally I've found 16k/16k to be the fastest on my setup. It might help some people out that are looking for top performance from their RAID array setup.

Posted: Fri Dec 06, 2002 4:41 pm
by Buub
Well, cluster size might be irrelevant to some people, because 4K is the max that some software will deal with (like some defragmenters, including the one I use: Raxco PerfectDisk).

But yeah, in general, more information is always welcome. :-)

Posted: Fri Dec 06, 2002 10:10 pm
by Chris Nolan
I have a question about the conclusion drawn on this page:
Transaction rate results really start to space out with four-drive RAID 0 arrays. First, we see that HighPoint's RocketRAID 133 is slower in every access pattern and at every load level. We're working with a four-drive array here, and the IOMeter tests are clearly saturating the RocketRAID 133's two IDE channels. In this case, two channels of ATA/133 aren't enough to keep up with a four-drive RAID 0 array.

How does this square with the throughput rates on this page which for the RocketRAID 133 are around 2MBps (though do reach 3MBps in the Web Server benchmark). How can 2MBps (or even 3MBps) overwhelm an ATA-133 channel?

Posted: Fri Dec 06, 2002 11:16 pm
by Chris Nolan
I'd like to suggest an answer to my own question.

Isn't the real cause of the lower performance of the RocketRAID 133 that the ATA buss protocol has no disconnection mechanism. A transaction once started to one buss target must complete before the next can begin; the buss protocol enforces serialisation and prevents HDDs on the same buss from operating in parallel.

Posted: Sat Dec 07, 2002 5:27 pm
by Mr Bill
Dissonance wrote:
I can pull together some screen shots and email them to you if you'd like, Mr. Bill. Otherwise I can try to rig up my server to host them and just post links here. I'm assuming that a screen shot of each configuration tab for IOMeter should suffice?

Also, you might not want to run the tests on a drive that you're currently using, since running the test involves 'preparing' the drives, which pretty much means filling them to capacity.
Either E-mail or posting links would be fine. A screen shot of each tab would be perfect. Since others might want to try the test also, maybe posting links would be best. I've used IOMeter for the 64K streaming tests and I have plenty of disk space on all drives involved. Afterward I just delete the test file.

Posted: Tue Dec 10, 2002 10:31 pm
by the_silver_bullet24
you should do the ATTO Transfer rates - Read and Write and the IOmeter tests RAID vs SCSI. that would be interesting but don't use the Adaptec or the fasttrack because that would be a waste of time. RAID 1 and 10/0+1

Posted: Wed Dec 11, 2002 9:42 am
by Mr Bill
2CPU has a thread about TR's IDE raid review. I've benchmarked a software raid 0 using IOMeter here:
http://forums.2cpu.com/showthread.php?s ... adid=29098
Its an all-in-one figure for easy comparison with the Tech Report Review.

Posted: Wed Dec 11, 2002 7:23 pm
by Goblyn
How have you gotten IOmeter to report non negative numbers on systems that have 2GHz processors or higher? Is there a patch for the windows version to correct this?

If anyone has any information about this I would be very grateful.

-G