WTF - See if this makes any sense to you...

[EricMartello]

I have a really weird problem with my RAID 0 array, at least as far as ATTO can show. My write speeds absurdly high, but write speeds are ridiculously low! Both should be around 100 MB/s, and that's what I get with Sisoft 2005's bench. I wanna know why my ATTO score looks like this:

http://www.htpcforums.com/index.php?showtopic=485

 

[EricMartello]

Anybody?

 

[stevty2889]

Why not run HDtach, sandra is not a good hard drive benchmark, and I don't know that other program, but if you run HDtach, I have something to compare it to.

 

[Madwand1]

4 MB is too small a test size -- with HD write-caching on, I think the HD is saying "yep, I wrote that, carry on!", immediately, and artifically boosting your reported write speeds.

You're looking at your buffer transfer speeds with these tests, and with SATA 3.0 Gb/s, these look nice.

With 16 MB cache drives at 2x RAID 0, even 32 MB write tests could be too small.

And maybe while the read test is being done, the write buffer is being flushed, thereby degrading that number.

It looks like a bad benchmark.

 

[EricMartello]

Yeah i figured the benchmark was skewed somehow. I usually trust ATTO results; I thought it was supposed to bypass caching. I'm pretty sure my reads are not as bad as the benchmark makes them seem.

I gotta say tho, benchmarks aside, this disk config is probably the best I've had on my desktop. Very responsive, and by setting the system to focus on "caching" in device manager, as well as disabling last access writes, it really hauls.

 

[EricMartello]

Just added HDTach to that collection. Sandra's disk bench is actually not that bad, I'd say it is providing the most accurate results of the three...but even so it's still off.

http://www.htpcforums.com/index.php?showtopic=485

 

[SickBeast]

Check how long it takes to load Far Cry levels, then compare to the AT review of the new 150GB Raptor.

I would actually be interested to see your results as I could very easily go to RAID-0 if it's worth it (I have a 250GB WD 16MB drive).

 

[obeseotron]

My personal experience has been no subjectively noticable impact on game load times, moderately improved boot time, and greatly improved speed for manipulating large files (usually video editing). This is on a pair of Diamondmax 10 300GB drives, which according to anand's benchmarks is one of the faster 7200RPM SATA drives, the raptor may behave differently. If you intend to boot off any array, prepare to learn all about slipstreaming drivers...

 

[Googer]

Originally posted by: stevty2889
Why not run HDtach, sandra is not a good hard drive benchmark, and I don't know that other program, but if you run HDtach, I have something to compare it to.

IOmeter and PCpitstop.com

 

[Googer]

Originally posted by: EricMartello
My write speeds absurdly high, but write speeds are ridiculously low!

Which one is it?

 

[aigomorla]

i believe that drive is a NCQ drive. In raid NCQ drives have a advantage over non NCQ drives. I was actually thinking of setting up a 1.2 TB raid5 partition with those drives and replacing my 3 74gig raptors with 2 150 gig versions. but now that i see the results, im getting mighty tempted to do that. :X

 

[stevty2889]

Thats with just 2 drives?? Your sustained rate is more than double the 101.4mb/s I get with my 160gb 8mb cache 3G drives, with your burst rate ~180mb/s faster(mine is 286mb/s). Plus your seek time is less than half of mine too. Those are some fast drives!

 

[Bobthelost]

Originally posted by: aigomorla
i believe that drive is a NCQ drive. In raid NCQ drives have a advantage over non NCQ drives. I was actually thinking of setting up a 1.2 TB raid5 partition with those drives and replacing my 3 74gig raptors with 2 150 gig versions. but now that i see the results, im getting mighty tempted to do that. :X

Are you sure? The advantage from NCQ is more to do with randomised acces paterns found in server enviroments and as far as i know the physical distribution of data on the platter will be more localised not less with RAID0. As such RAID0 should suffer just as much of a penalty as a single drive.

 

[EricMartello]

Yeah the Maxline III drives are pretty good. They're actually designed for server use and have relatively high IO/s when compared to typical desktop SATA drives. I dont seem to have that data corruption issue that has been mentioned to be present with maxtor drives and the nforce4 chipset...tho now that I said that, I will probably have a corrupted disk tomorrow. NCQ seems to have mixed results depending on the particular drive, sometimes it is faster and sometimes it actually reduces performance. With the Maxline III drives NCQ does better it on rather than off, probably because they were designed for server environments.

I think the major benefit of raid 0 on a desktop is working with large files; that includes your system's page file, which is a great reason to use raid 0 on a desktop. It also reduces load times on some games when properly configured. Remember that stripe size has a big impact on the array's performance. It's not about matching the stripe to the cluster size, it's about selecting a stripe that offers the best match for the size of the data files you will use the most often. I used these drives in standalone mode for over a year, and now using them in raid 0 I can say there is a definite and noticeable improvement in overall system responsiveness using the default 64KB stripe of the nforce4 bios.

As for that myth about increasing the chance that you will lose your data and/or reduce the effective MTBF by using raid 0, that is a complete farce and I see a lot of people spreading it here like its fact. If all drives in the array have the same 1,000,000 hour MTBF as mine do, the array has an effective 1,000,000 MTBF with the same probability of failure as a single drive. It is no more or less redundant than a single drive. Anyone who is worried about the integrity of their data should be backing up anyway, and not hoping their disk subsystem never craps out.

 

[Bobthelost]

Hmm, the maxline III range does have roughly equal performance with and without NCQ (source = storagereview), better for some things and worse for others, but in total ah heck all difference. Interesting, of course NCQ does make a massive difference for server type roles, but that's irrelevant here. Could just be that the maxline III drives are pretty crap all round.

The data loss risk is not a myth at all. It's common sense backed up with mathematics. A 2 disc RAID0 array has two components which can die at any one time, causing the loss of all the data rather than just half of it. But it's also rather irrelevant if you make regular backups.

 

[EricMartello]

Originally posted by: Bobthelost
Hmm, the maxline III range does have roughly equal performance with and without NCQ (source = storagereview), better for some things and worse for others, but in total ah heck all difference. Interesting, of course NCQ does make a massive difference for server type roles, but that's irrelevant here. Could just be that the maxline III drives are pretty crap all round.

NCQ is great when the drive access patterns consist of frequent I/Os...in which case reordering the drive's commands would provide better performance. On the desktop, most reads are sequencial and NCQ's reordering of the commands just slows it down by adding an unnecessary step.

The data loss risk is not a myth at all. It's common sense backed up with mathematics. A 2 disc RAID0 array has two components which can die at any one time, causing the loss of all the data rather than just half of it. But it's also rather irrelevant if you make regular backups.

Well that's some bad math...in a raid 0 array, if one drive dies that the data is essentially unrecoverable since 1/2 of it has been striped onto the other drive. The same is true if all the data was on one drive.

The act of creating a raid 0 array of 2 or more drives does not increase the individual drive's probability of failure. So whether you have 2 drives or 5 in a raid 0 configuration, the probability of you losing all data due to disk failure is the same as if you had 1 in a non-raid configuration.

 

[Madwand1]

Originally posted by: EricMartello
The act of creating a raid 0 array of 2 or more drives does not increase the individual drive's probability of failure. So whether you have 2 drives or 5 in a raid 0 configuration, the probability of you losing all data due to disk failure is the same as if you had 1 in a non-raid configuration.

The first part, "does not increase the individual drive's probability of failure" is correct, the conclusion is not, because any single failure is a system failure.

System MTBF for n components is calculated as follows:

1 / MTBF (system) = 1 / MTBF (1) + 1 / MTBF (2) + ... + 1 / MTBF (n)

For n devices with equal MTBF, this simplifies to:

MTBF (system) = MTBF (1) / n

So 2 x RAID 0 has 1/2 the system MTBF of each drive, not counting the controller.

http://www.relexsoftware.com/customers/cs/versatel.asp
http://www.storagereview.com/guide2000/ref/hdd/perf/raid/concepts/relRel.html

However, if we assume that the published MTTF figures (1,000,000 hours for Maxline III), then the net results aren't so significant -- for a 2 drives, in 1 year I calculate 1% increase chance of failure; in 10 years, 8%. Assuming 1/2 that MTTF, I get 2% and 12% respectively. Assuming 1/2 the Maxline MTTF, and a 4 drive array, I get 5% in the 1st year, 21% in 5 years, and 34% in 10 years.

But this is really off-topic in this thread -- we still don't know why the OP's getting crazy performance numbers here. They are very high, and probably indicate problems with the benchmarks. This could matter, because people often use the same benchmarks elsewhere.

I tried 2 current Maxtor DiamondMax 10 300 GB / 16 MB cache SATA II drives using NForce 430, and didn't get anywhere near those transfer rates, nor those seek times. These drives are also supposed to have NCQ, though perhaps it's not the same tuning / sophistication of the Maxline III.

I don't know the reasons. I think it's not good to just believe those numbers. Simple transfer tests might give more reliable figures. E.g time to copy files from memory to drive -- via applications or via file copies where the source is known to be cached. If the seek times are really as great as the benchmark tests say, then copying to & from the same array should also give very good transfer rates.

I find Sysinternal's CacheSet useful for such tests. (It can clear the cache, thereby eliminating the effect of the cache on repeated tests.)