[692] in Info-AFS_Redistribution
benchmark comparisons of UNIX/NFS/AFS
daemon@ATHENA.MIT.EDU (BARTON@dalnessc.vnet.ibm.com)
Tue Mar 31 21:06:45 1992
Date: Tue, 31 Mar 92 16:12:53 CST
From: BARTON@dalnessc.vnet.ibm.com
To: info-afs@transarc.com
We are evaluating switching to AFS for our clustering also. It's really
difficult to compare the performance of AFS and NFS because each has
benefits. Here are a couple things that come to mind.
1) MTU size. I noticed that AFS packets are limited to 1480 bytes
regardless of the network type.
This has a dramatic impact on performance. I just did a comparison
on 16Meg Token Ring and this is what I got.
MTU size NFS Read rate
1480 479KB/sec
4096 1289KB/sec
AFS gives me about 500KB/sec no matter what MTU size. It leads me to
believe that AFS would be just as fast if it used bigger packet sizes.
Is there any way I can set this parameter?
FDDI numbers are much higher but still show the same degradation when
using small MTU sizes. It becomes more obvious when you go to
SLA where you can use even larger MTUs.
Also, server CPU loads tend to go up as MTUs go down.
P.S. I realize that this was done because of all the ethernets in the world.
2) Some of our clustered applications use NFS because the data is too big
to fit on the disks. It seems like the AFS cache might cause problems
here. What if you have to write out a 1 Gigabyte file?
Robbie Barton
*** These numbers are on completely untuned systems and should not be
*** considered best case or completely accurate.
