[OpenAFS] Perfomance

[Charles Clancy]

> Between a P-200-server and a K6/2-350-client (cachesize: 100MB ) I got 600 kByte/s.
> After I made a 4MB-Ramdisk for the client's cache-manager, it increased to about
> 3MByte/s.

3MBps is pretty good.  I get around 1.5MBps when copying to localhost.
My server/client is an UltraSparc-500 MHz.  The second time I tried
copying the file, I get around 2.7MBps, which seems to imply some sort of
limit on how fast the cache manager can deliver a cached file.

Other benchmarks: 4.6MBps between partitions of a normal local UNIX
filesystem.  40MBps from local to /tmp.

> All PCs are non-high-end (IDE-HDDs, Cheap NICs) but 3MByte/s is very
> slow (within a switched 100MBit-Ethernet). FTP sometimes reaches the
> 10MByte/s-mark in the same environment.

Are we talking 3200 RPM IDE or 7200 RPM IDE?  If it's the former, that's
probably your slowdown.  See how fast you can copy between partitions
locally.  Having a 100 Mbps switch does you no good when your IDE bus is
only running at 15 Mbps.

> I was told that the cachemanager could be the problem. Is there a chance
> to bybass it (at least for accessing the local subnet)?

We've had this discussion before.  First of all, it would involve major
changes to the way the AFS client works.  Second, you'd have cache
consistancy problems with other AFS clients if one client didn't have a
cache.

[  t charles clancy  ]-[  tclancy@uiuc.edu  ]-[  uiuc.edu/~tclancy  ]