[Brian]

Last week Western Digital released the new 600GB VelociRaptor, which just like its predecessor, took the spot of fastest consumer hard disk drive. With the single hard drive not being able to keep up with average SSDs, we decided to take a look at multidrive performance. In this review we see how well the Western Digital VelociRaptor 600GB performs in RAID 0 and RAID 1 configurations using the 6Gb/s LSI MegaRAID 9260-8i controller card. The results are impressive.


Full Story

[geshel]

Hi Brian,

What is the stripe size you used for the RAID0?

One interesting thing to note is how the RAID1 outperformed the RAID0 in the IOMeter "Web Server" test. Whereas in all of the low-level tests, RAID0 outperformed RAID1.

The reason for this is striping. It indicates that the Web Server IOMeter test issued a significant number of reads that were larger than the stripe size; each of these reads caused a seek on both HDDs for the RAID0. Whereas for the RAID1, each read could be handled by a single drive, leaving the other free.

I don't remember the spread of IO sizes that the Web Server test uses, however it's likely that you'd see the difference between the RAID0 and RAID1 configurations change in that test as you changed the stripe size of the RAID0.

[Brian]

You know, we had the stripe size in the review but it must have been cut by accident

64k stripes for both.

[geshel]

You should try the benchmarks again with a larger stripe size. My estimates show that a 256KB stripe size would improve the random read performance of the RAID0 by between 10% and 50% for reads between 8KB and 2MB. A stripe size of 1024KB would improve things 20-75%.

Write performance would increase as well, though I don't have numbers for that. I'm pretty sure that the IOMeter results would show the RAID0 handily outpacing the RAID1 for either 256K or 1024K stripe size.

[BuddyLite]

geshel, on 15 April 2010 - 03:20 PM, said:

You should try the benchmarks again with a larger stripe size. My estimates show that a 256KB stripe size would improve the random read performance of the RAID0 by between 10% and 50% for reads between 8KB and 2MB. A stripe size of 1024KB would improve things 20-75%.

Write performance would increase as well, though I don't have numbers for that. I'm pretty sure that the IOMeter results would show the RAID0 handily outpacing the RAID1 for either 256K or 1024K stripe size.

Come on guys ! Almost any midrange SSD will trounce mechanical drives. WD has built a dinosaur.

[Brian]

Dinosaur? I don't think so, there's still a place for performance spindles. That market may not be very large, or growing even, but given the reads we had on the initial review, there's clearly still interest.

[geshel]

HDD performance and failure modes are very well-known quantities. SSDs are not. My organization, for one, is not going to put any business-critical assets on SSDs until that changes.

[kittle]

Yes SSDs are much faster - but thats not the issue.

its reliability. SSDs havent been around long enough to establish if they are reliable or not. Also what happens when they fail? do you get an expensive paperweight? Some do, some dont. And there is also the lifetime issue of SSDs

With a mechanical drive, its a known quantity. If it dies -- I can try recovery software, a controller swap, and even dedicated data recovery services.

Once SSDs reach the "known quantity" state, they will be more widely accepted.

[Brian]

We had some time between reviews to go back and revisit the VelociRaptors and modify the RAID 0 stripe size.

As you might remember with our 64KB stripe we had the following results with the HTPC and Productivity traces:

HTPC 64K: 111.51MB/s, 2432I/Os, and 3.241ms
Average 64K: 52.89MB/s, 1794I/Os, and 4.442ms

The first new RAID stripe size we aimed high with 512K which improved the HTPC test with a higher concentration of large transfers, but lowered the Productivity test that had higher random activity.

HTPC 512K: 120.09MB/s, 2617I/Os, and 2.993ms
Average 512K: 52.75MB/s, 1786I/Os, and 4.462ms

Next we toned it down and went with 256K. HTPC barely changed from 512K (easily within the range of error) but our Productivity test did get a healthy boost this time around.

HTPC 256K: 120.05MB/s, 2616I/Os, and 3.007ms
Average 256K: 55.42MB/s, 1876I/Os, and 4.233ms

We were actually hoping for the huge gains geshel was mentioning but that might only have been possible with a consistent synthetic benchmark. With the traces they have such a broad mix of varying user activity that it looks like the smaller stripe size offers more benefits in both situations. Although the argument could be made to bump it up slightly from 64K to 256K, but not any higher than that.

[fqv]

I don't know if you are aware of this but you can actually use SSDs and mechanical drives on the same system. Most motherboards come equipped with at least a couple of SATA ports, you simply connect the SSD to one of them and the mechanical drive to one of the others. This way you can have a SSD where you need quality (operating system and software), and HDD where you need quantity (games, music, movies, etc). Even though SSDs are more expensive per GB you do not actually need that much, and the total price of the mixed set may actually come out lower than the VelociRaptor(s).