[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Software raid 5 performance tuning?



hi ya marc

On Sat, 15 Oct 2005, Marc Dreher wrote:

> after a recent (data) lossy HD crash I decided to build a software raid
> 5 (3 disks)with Debian Sarge (2.6.13 Kernel) to keep that from
> happening again :-)

3 disks in raid5 means 1/3 of the total disk space is not usable

> Works pretty well but I am not really satsfied with the write
> performance (via samba).

writing data into raid5 will be slow, since it's writing parts of the same
data n-times to each disks 

reading should be faster ... but not always

> When uploading files via samba I get around 6 MB per second. Uploading
> to a non-raid disk works at full 100Mbit network throughput (10 MB
> p/s).

presumably the target disks getting the uploaded files is on the same
pc ???  and the obsvervation you're seeing is normal/typical

> I have one ATA disk as master on each of the two onboard IDE channels
> and the third SATA disk on one of my two SATA ports.

odd combo but good to have only 1 disk on each ide cable

> I used a chunk size of 64KB with no strides.

more parameters to play with 
 
> So before having to experiment with all kind of settings I hope that
> somebody does have any tips or tuning hints for me?

increasing write performance is not ez

	- you probably can change your nfs read/write transfer sizes
	to get some improvements

	- you probably can get "faster" transfers if you compress it
	before sending it

	tar zcf - stuff | ( cd /mnt/samba ; tar zxfp - )

> I think with current hard- and software I should at least be able to get
> enough write performance to fill up my 100MBit network while uploading :-)

8bits * 10MByte/sec --> 80Mbps .. that's pretty good 
anything say aroun 70% - 90% of the rated networkwork thruput is good

you will need to go to GigE network

8bits * 100MB/sec --> 800Mbps ... good enough for gigE entwork
and the disks are rated ( by marketing folks ) at 100MB/sec (120MB/sec)
	or faster and your new bottleneck will be the disk rotations

c ya
alvin




Reply to: