On Wed, Nov 07, 2012 at 02:57:46PM +0100, Edgar Fu? wrote: > > The winner was 64k fsbsize on 128SpSU (resulting in one fs block per stripe > unit and four per stripe).
This looks wrong. The 6.0 kernel will never, under any circumstances, issue a transfer larger than 64K to a disk device, not even a pseudodevice like RAIDframe -- so this configuration guarantees that you will never be able to do a full-stripe write, so that every write becomes a read/modify/write cycle, tripling the I/O load. What fragment size are you using? Use of 64K blocks is quite rare because the kernel already clusters I/O to 64K almost all the time for both reads and writes, regardless of the blocksize. Thor
