On Wed, Mar 08, 2000 at 06:52:52PM -0000, Matthew Clark wrote:
> Hmm.. well you may think 26Mb/Sec is poor for writing.. I would be drooling
> at such vast speeds..
>
> Would you mind telling me how you set up your raid array (i.e. policies) and
> filesystem (inodes, block sizes, strides etc)...I'm seeing 2M/b per sec on
> writes and only 16-17mb/sec on reads.. sequential or random!!!
>
I believe for the 26MB/s across two chains I set up the Mylex board to
have 64KB stripe and 64KB segment size. For ext2, I modified it to
use 8KB block size (which is the page size on Alpha, Ted says that
won't work on Intel). So I called mke2fs like this:
mke2fs -b 8192 -R stride=8 -i 16384 -s 1 /dev/rd/c0d0p1
I also use write-back, which increased the performance a bit. The
current equipment I have is going back to Compaq in a few days but I
already have more loaners in (they need to be installed). I'll have 4
Mylex DAC1164P controllers and 40 10K RPM 18GB drives in 8 cabinets to
try to get as much performance out of as possible. My goal is 100MB/s
on a single node with a maximum of 3 controllers in one node. I
really doubt I'll get there as it seems that with the current
configuration and driver these controllers don't scale well. I was
getting around 19MB/s with one chain; adding the second chain only
gave me another 7MB/s more. The second chain should have doubled my
performance, this is why I am nearly certain that I'm hitting a
read-modify-write sequence on nearly every write.
I'm going to try to find a SCSI analyzer and hook that up to track
what is happening on the bus. When we got the big IBM machine in a
few years ago (which I wasn't here for), they did this and found they
were always getting a read-modify-write sequence for writes. Our
engineers here worked (bullied) IBM into modifying firmware code,
driver code, VFS layer code, and JFS/GPFS code to be able to support
block transfers in such a way that you are always dealing with stripe
sized chunks and never have to do read-modify-writes. The numbers I
was quoted was up to 10 times the performance once this was done.
BAPper
