Disk benchmarks

[Roy]

I recently have been charged with the task of analyzing our disk sub systems (we have SCSI and SATA RAID packs). I was confused by the results that Performance Test returned. Below are two configurations and the results returned:

Intel SE7505 Dual Xeon 2.66GHz
2GB RAM
3Ware Escalade 9000 SATA RAID (w/ 128MB cache)
WD1200JD 7200 RPM SATA disks (3 disks RAID 5, 1 hot spare)
Stripe size is 16KB
OS Windows 2003 Server (standard)

1024MB 100% Read 100% Sequential uncached 16KB block => 23.7MB/sec

1024MB 100% Read 100% Sequential cached 16KB block => 285.1MB /sec

1024MB 100% Write 100% Sequential uncached 16KB block => 17.2MB /sec


Intel STL2 Dual PIII 799MHz
1GB RAM
Adaptec 2100S SCSI RAID (w/ 64MB cache)
Seagate st31804 7200 RPM SCSI (SCA) disks (4 disks RAID 5, 1 hot spare)
Stripe size is 32KB
OS Windows 2000 Server

1024MB 100% Read 100% Sequential uncached 16KB block => 12.8MB/sec

1024MB 100% Read 100% Sequential cached 16KB block =>
12.6MB /sec

1024MB 100% Write 100% Sequential uncached 16KB block => 19.2MB /sec

The 3ware controller is rated at 400MB/Sec reads

The Adaptec controller is rated at 80MB/Sec reads

Any help understanding these numbers would be greatly appreciated.

Roy

[David (PassMark)]

> The 3ware controller is rated at 400MB/Sec
> The Adaptec controller is rated at 80MB/Sec

Yes, but you need to be aware that these numbers are just used for maketing and have very little relation to performance in the real world.

In the real world, the disk performance is normally limited by the disk drive hardware. RPM, data density and seek times are what counts. If you are doing tests with small files and / or non sequential files, then caching and command queuing also become important.

In your case, your 1st machine has more main ram (for caching), more cache on the controller and faster CPUs. It is no suprise that the cached read test looks very good. Almost all the data would have come from one cache or another.

-----
David

[Roy]

> The 3ware controller is rated at 400MB/Sec
> The Adaptec controller is rated at 80MB/Sec

Yes, but you need to be aware that these numbers are just used for maketing and have very little relation to performance in the real world.

In the real world, the disk performance is normally limited by the disk drive hardware. RPM, data density and seek times are what counts. If you are doing tests with small files and / or non sequential files, then caching and command queuing also become important.

In your case, your 1st machine has more main ram (for caching), more cache on the controller and faster CPUs. It is no suprise that the cached read test looks very good. Almost all the data would have come from one cache or another.

-----
David

Thank you for your reply David. Another question or two if I may.

When you developers are testing the disk test, do they have a "rule of thumb" on what to expect from a disk?

Is there a formula that they use for calculating antisipated throughput?

My goal is to detrumine what a good RAID-5 controller/disk combination should report to PerformaceTest.

Thank you again for you time and efforts. I truely do appreciate your products.

Roy

[David (PassMark)]

There is no real formula.

But good drives should be above 20MB/Sec for read and write.
(Uncached, large block size, defagmented disk, squential access, no background tasks)

With RAID-0, I would be expecting 30MB/Sec or better. Even 50MB/Sec - 100 MB/Sec are possible depending on your drives

Caching can push results to 250MB/Sec+ becuase the physical disk is not used.

How RAID affects the results depends on the stripe size and RAID level (0 - 6) and the controller itself. For example RAID-0 should be faster becuase both drives work in parallel to read / write the data. For RAID-3 the read speed can be good but the write speed poor (becuase of parity, error-detection).

-------
David