Topic: Hardware RAID performance?

[Dan Brown]

I recently purchased a "new" used server to replace my old one; the new server is a Dell Poweredge 4300 with their PERC 2/SC RAID controller (a rebranded AMI Megaraid, AIUI).  This controller supports Ultra2-Wide SCSI at 80MB/sec.  I've created a RAID 5 array of 4 36GB, 7200 rpm drives, all of which also support 80 MB/sec transfer rates.  I'm kind of surprised at the performance of this system--I'd expected that with multiple spindles, fairly fast drives, and a fast interface, I should be seeing pretty quick transfer rates.  However, this is what I'm seeing:

[root@dell root]# hdparm -tT /dev/sda

/dev/sda:
 Timing buffer-cache reads:   128 MB in  1.00 seconds =128.00 MB/sec
 Timing buffered disk reads:  64 MB in  2.32 seconds = 27.59 MB/sec

That's the highest I've seen yet; more commonly the disk reads run around 15 MB/sec.  My old server has a pair of 80 GB UDMA 100 drives in a software RAID 1 configuration, on an old motherboard that only supports UDMA 33.  Here's what I get on that:

[root@e-smith mp3]# hdparm -tT /dev/hda

/dev/hda:
 Timing buffer-cache reads:   128 MB in  1.76 seconds = 72.73 MB/sec
 Timing buffered disk reads:  64 MB in  2.35 seconds = 27.23 MB/sec

Am I wrong in thinking I should be seeing significantly better performance than this?  Assuming that I'm right, what should I look at to try to correct this?  I'm running SME 5.6 on the server at the moment, but I had the same results with 6.0b3.  I expect I'll need to check in with the Dell forums, but thought I'd check here first.  Any suggestions on this would be greatly appreciated!

[Boris]

27 MB/sec sounds about right for the configuration you've described.
You will not see much improvement on the test like this. The real benefits you will see if it serves multiple requests simultaneously. Slow down threshold should be higher and CPU utilization will be lower.

Here some benchmarks for your reference:

DELL 2550 4x18 U80 10000 drives (RAID5)
/dev/sda:
 Timing buffer-cache reads:   128 MB in  0.59 seconds =216.95 MB/sec
 Timing buffered disk reads:  64 MB in  2.48 seconds = 25.81 MB/sec


The highest performance I’ve seen (DELL 1650, SCSI U160 2x36GB, 10000 RPM drives mirror)
/dev/sda:
 Timing buffer-cache reads:   128 MB in  0.47 seconds =272.34 MB/sec
 Timing buffered disk reads:  64 MB in  0.96 seconds = 66.67 MB/sec



Typically on IDE (U100) you see something like +/- 40 MB/sec

/dev/hda:
 Timing buffer-cache reads:   128 MB in  1.11 seconds =115.32 MB/sec
 Timing buffered disk reads:  64 MB in  1.65 seconds = 38.79 MB/sec

/dev/hdc:
 Timing buffer-cache reads:   128 MB in  1.14 seconds =112.28 MB/sec
 Timing buffered disk reads:  64 MB in  1.40 seconds = 45.71 MB/sec

[Dan Brown]

Interesting.  I'd seen reviews of the 3Ware cards, where the RAID0 performance increased nearly linearly with the number of drives--I'd expected I'd see something similar with RAID5.  Maybe not.  Thanks for the info!

[Kelvin]

Hi Dan,

>RAID0 performance increased nearly linearly with the number of drives--I'd
>expected I'd see something similar with RAID5

Unfortunately, RAID 5 calculations is substantially more complex than RAID 0 hence the performance penalty is also higher. RAID 0 has no fault tolerance while RAID 5 needs to compute parity information for fault tolerance.

Aside from just the task of recombining / splitting data coming from and going to the drives, you must remember the more drives you have, the more the controller needs to synchronise reads / writes etc.

Kelvin

[Shad Lords]

I've actually got a 3ware 7500-12 with 3x160Gb in a raid-5 configuration.  The following are some numbers that I obtained.

With the system loaded.  Doing a virus scan and backup at the same time:

 Timing buffer-cache reads:   128 MB in  0.85 seconds =150.59 MB/sec
 Timing buffered disk reads:  64 MB in  1.07 seconds = 59.81 MB/sec

With the system idle:

 Timing buffer-cache reads:   128 MB in  0.76 seconds =168.42 MB/sec
 Timing buffered disk reads:  64 MB in  0.88 seconds = 72.73 MB/sec

-Shad