Re: [Xen-users] Network based storage - NBD/AoE/iSCSI other?

[Luke Crawford <lsc@xxxxxxxxx>]

On Mon, 2 Oct 2006, John Madden wrote:
> (Saturating the link is fine as long as your performance is good as
> well.)  Anyway, the numbers I've read from them show something in the
> neighborhood of really terrible:
> http://www.linuxjournal.com/article/8149 -- 23.58MB/s reads.  I'd expect
> that performance out of a single ATA disk. IOPS is obviously something
> different, but I don't believe that article mentions it.
I find disk benchmarks with dd amusing.  I soppose it works 
if all you are using it for is backup and other sequential tasks-  but for 
my load, you'd want to benchmark 20 parallel dd processes, and then 
latency is more important than throuput.
The funny thing is, IDE does usually win the thruput race, if you 
stripe/raid over a couple busses-  23MB/sec is a lot less than I would 
expect in a sequential transfer from 10+ ide disks.
Still, we saw the same thing on our (horribly overpriced)  EMC NAS at 
one of the the last places I worked-  I think the partition we were 
testing had 4 10KRPM scsi drives, and it had 4GB of cache (in a redundant 
configuration, running in write-back mode) and it had something like 6 
gigE connections.  from one box, we got mediocre sequential speed;  but 
we found that we could do the same transfer from 10 seperate client 
boxes, and still get the same speed from each client-  making it not so 
mediocre after all.
(I would be quite suprised if the corraid box could stack up to the $100K+ 
EMC, even given the AoE advantage. (the EMC was NFS)  but my point is that 
being able to adaquately service multiple clients is usually more 
important than sequential transfer to just one)
oh, also;  Jumbo Frames make more difference than the NFS server used for 
large transfers over NFS;  I do not know how true that is of AoE, but if I 
were troubleshooting slow AoE access, jumbo frames would be one of the 
first things I'd try.

_______________________________________________
Xen-users mailing list
Xen-users@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-users