I recently bought expensive, expensive WD Raptors for my new system, hoping for some real speed. So today I ran some hdparm benchmarks and the results were... surprising seeing that reading the disks directly was almost twice as fast as reading from RAID.

sdd and sdc are the Raptors; partition for md0 lives on 5.

/dev/sdc5:
Timing cached reads: 5314 MB in 2.00 seconds = 2660.02 MB/sec
Timing buffered disk reads: 242 MB in 3.02 seconds = 80.09 MB/sec

/dev/sdd5:
Timing cached reads: 5650 MB in 2.00 seconds = 2827.28 MB/sec
Timing buffered disk reads: 234 MB in 3.02 seconds = 77.61 MB/sec

/dev/md0:
Timing cached reads: 5754 MB in 2.00 seconds = 2879.53 MB/sec
Timing buffered disk reads: 142 MB in 3.00 seconds = 47.27 MB/sec

I repeated the test many times and the results were all similar, except that md0 occasionally get 55M/s. This was very shocking compared to my old PATA system with 7200rpm hdd:

/dev/hda:
Timing cached reads: 244 MB in 2.01 seconds = 121.37 MB/sec
Timing buffered disk reads: 144 MB in 3.01 seconds = 47.88 MB/sec

Of course, I realize hdparm isn't a very accurate benchmark, so I ran dd to measure read/write. Those results were slightly more consistent:

Reading from md0, writing to md0
root@plato:~# dd
if=/dev/md0 of=/tmp/dd_md0
1956353+0 records in
1956352+0 records out
1001652224 bytes (1.0 GB) copied, 21.1957 s, 47.3 MB/s

Reading from individual disk, writing to individual disk
root@plato:~# dd if=/dev/sdc of=/tmp/dd_sdc
1116129+0 records in
1116128+0 records out
571457536 bytes (571 MB) copied, 17.722 s, 32.2 MB/s

Reading from individual disk, writroot@plato:/boot# dd if=/dev/sdc of=/boot/md_sdc
dd: writing to `/boot/md_sdc': No space left on device
141369+0 records in
141368+0 records out
72380416 bytes (72 MB) copied, 3.51089 s, 20.6 MB/s
ing to individual disk:

This seems to make sense, but every number except for the 47M/s is disappointing and others report a much higher throughput (albeit with Windows tools, which may or may not be comparable). Nevertheless, they certainly *feel* fast, especially with NCQ and AHCI that make them feel like true SCSI drives; latency is low and huge disk access does not freeze the system, so I guess it's fast in the areas that matter.

Anybody have any idea what could be causing these strange results?

Software RAID 0 is probably fairly slow in general. Also, the drive wasn't mounted during any of the tests, was it? If so, it can obviously be affected by that.

Well, md0 was mounted, that probably threw a monkey wrench into hdparm. I have a fast processor (Core 2 Duo E6750), so can software RAID0 really have that much impact? If so, then is there no point in using software RAID0? ICH9's "hardware RAID" is still powered by the CPU in Windows.