Software RAID5 - poor write performance & freezing
 

I've been playing with the software RAID5 abilities of the 2.6.13 kernel. After numerous tests, I've settled on a 128k stripe setup on 4 250 GB drives.

After the array is created and it's synced, I get really poor write performance.

About 8MB/Sec, which is 25% or less of a single drive. I'm not looking for stellar performance, but something akin to a single drive is what I was expecting.

In addition to this, when writing to the drive, the system will appear to buffer a ton of data, then dump it to the raid array all at once, halting all other IO on the drive until the write is done. This causes a HUGE problem, since I need to be able to write and read from the drive at the same time in modest amounts. I can easily do what I need with a single disc... why isn't this working with the RAID5 setup?

So in short:

How can I increase the write performance of Software RAID 5 in linux?
How can I smooth out the writing to the drive so it doesn't buffer and then dump the entire contents of the buffer to the drive, halting all other IO on the drive?

Thanks!

Im getting;

/dev/md1:
Timing cached reads: 1904 MB in 2.00 seconds = 949.77 MB/sec
Timing buffered disk reads: 150 MB in 3.04 seconds = 49.35 MB/sec

with 4 raptors....the only difference between Raid 5 and no raid with these drives is my read levels go down to 12.....cached reads stay at 9hundred-sumthn'.

/dev/md1:
readonly = 0 (off)
readahead = 256 (on)
geometry = 15648/2/4, sectors = 97642752, start = 0

/dev/md2:
Timing cached reads: 1900 MB in 2.00 seconds = 948.72 MB/sec
Timing buffered disk reads: 160 MB in 3.04 seconds = 52.71 MB/sec

Not sure how to improve performance on r5 other than stripe size.

Not haveing experianced this,I can not answer...sorry.

I think 128 kilobyte strip is huge for a single processor system to calculate the parity information for RAID 5 array. A dual or multi-processor system is required for softare RAID 5 setup. You can setup two RAID 0 arrays. When you format /dev/md0, make /dev/md1 be your journal drive. You get reduntancy and you also you get an increase write throughput, but at a cost of 250 GB less space. Also the read speed of RAID 0 is better than RAID 5 because of lower latency and slightly higher data throughput. You should leave RAID 5 to hardware RAID controllers like 3ware.

Well, that sounds reasonable... I will try playing with the chunk size and lowering it. The problem is, most of the data on this drive is going to be between 3 and 9 GB per file, so I went with a larger stripe size.

Regardless... several people I've talked to say they have no problems getting write speeds in the 20's on a PIII/1Ghz system with software raid 5.

I set this array up as a giant RAID-0 array (1TB) and got about 20MB/sec write speed sustained. That's about where I'd expect a raid-5 to be, or at least where I'd want it to be for my application.

Running this in a RAID-1 setup and losing 500GB of space is not really an option, since space is and controller ports are at a premium.