Hello.

I am currently working on finding out if I can maintain a write speed of >160mB/s using 4x500GB harddisks under Linux.

The drives are connected to an onboard nvidia controller and
the raid0 is done using software (md). I'm using software raid0 because linux doesn't detect the raid if I use hardware (it just detects the 4 disks individually).
I am currently using ext2 since it has no journaling.
The chunk size of the raid is 128kB and the ext2 block size is 4096kb.

The actual files that are to be stored will be, on average, about 50gb.

I currently use dd, writing a 1.8TB file, to test the setup.

Currently I can maintain a speed over 160 mB/s until after 1.3TB (out of 1.8TB) has been used on the disk. After 1.4TB the speed slowly drops to 130mB/s (due to the nature of the harddisks of course). I do however
see a lot of spikes (jumping from 170 to 50 for example) in the writespeed, which I would very much like to avoid.

My question is this:

What raid chunk size should I use?
What filesystem should I use for max speed?
Is there anything I can do to tweak the system? (I'm only runnning processes I need).

Stability is not a problem, only speed is

Thanks in advance for any advice you might have.

Obviously you want to check that you are using DMA on all the drives with hdparm. Also, are you looking at lots of small files or more towards big files? This significantly impacts filesystem performance, so it's important to consider. Also, for what do you need >160 mb/s performance?

The disks are all SATA disks, I think DMA is on by default for those?
hdparm does not even work on them;
hdparm -i /dev/sda
/dev/sda:
HDIO_GET_IDENTITY failed: Inappropriate ioctl for device

The files will be large, around 40-50GB.
So far I have only used ext2 because it doesn't use journalling, I figured this would be the best
speedwise - I might be wrong.

I can't really say excatly why I need the write speed, it's a company secret and
I'm not allowed to talk about it - sorry.

One solution would be if I could tell md to use two disks starting from the outside of the
harddisk plates and two disks starting from the inside.
That way I would have around 160mB/s stable instead of going from ~340mB/s to ~120mB/s.

Any advice is welcome and appreciated

I would try using XFS rather than ext2. It has much better performance for very large files, such as the ones you described.