Subject: Re: RaidFrame poor performance
To: Jochen Keil <J.Keil@gmx.de>
From: Greg Oster <firstname.lastname@example.org>
Date: 01/20/2005 09:08:02
Jochen Keil writes:
> Daniel Carosone wrote:
> > To resolve this, if you can, add a fifth disk and choose your stripe
> > size and filesystem block size carefully so the filesystem blocks
> > exactly map to a full stripe write. Otherwise, you'd be much better
> > off with stripe+mirror over the four disks.
> My setup is quite similar to Mihai's. The differences are that it has
> one channel for every disk and i'm using 200gb disks.
> I'm satisfied with my performance (~11mb/s on random write) but after
> your email i was wondering if there could be some optimisation.
> At the moment i'm doing a benchmark on my 4 disk R5 setup and later i
> will do so on a 3 disk R5. What i'm wondering about is your comment
> about matching stripe size and file system block size.
> Default FS block size is 16kb for partitions above 1024MB. I'm using
> this R5 Layout:
> START layout
> # sectPerSU SUsPerParityUnit SUsPerReconUnit RAID_level
> 32 1 1 5
> I learned that 32 blocks equal to 16KB. As far as i can see my block
> size matches my stripe size so this is the optimum.
No. In this case, with "4 data disks", a full stripe will hold
4*32=128 blocks (64K) of data. (That "32" is per-component.)
So a block size of 64K will probably perform the best with this