On Thu, Mar 02, 2000 at 07:39:16PM +0100, Andrea Arcangeli wrote:
> On Fri, 3 Mar 2000, Brian Pomerantz wrote:
>
> >be wonderful if I could get a 64KB block size working. I could easily
> >use a single block per stripe with that.
>
> I think to stripe well you want the stripe size to be low, not the
> blocksize to be large.
>
> Here with LVM I use a 4k blocksize and a 4k stripsize.
>
> I don't think using a large stripsize make any sense. The hardblocksize is
> still 512 bytes. The most finegrined the stripsize is, the better the
> disks will scale. So I suggest to use 4k stripsize with ext2.
>
Unfortunately, the smallest stripe size I could get on the Mylex card
is 8KB. When I tried that size, I was getting half the write
throughput compared to a 64KB stripe size. Also, to really have
optimal performance I need to use at least an 8KB block size (to match
the Alpha page size). After looking at many of the current hard
drives, I noticed that the physical block size is variable depending
on where you are on the platter. Almost all of the drives I looked up
had a maximum transfer size of 64KB for a single transaction. I'm
wondering if making the stripe size equal this transfer size * data
drives would actually be the optimal configuration.
I think the consensus here at LLNL is that for the scale that we are
working with, a larger block size is better. We have also found that
if you can match the stripe size to the block size in a RAID 5, your
write performance will be at least twice as good compared to not
matching these up. On the ASCI Blue system, we saw a factor of 10
performance increase on the RAID 5 systems when we matched up the
block size to the stripe size. The reason behind this is you don't
have to worry about the RAID controller doing a read-modify-write of a
stripe because your block is the same size as the stripe.
In the end, I'll be using a clustering file system and not ext2. I
would like to use ext2 to play with RAID configuration and see where I
can get the best performance out of it, though.
BAPper
