[OpenAFS] Dreadful metadata performance

[Jamie Walker]

Since the start of the academic year, we've been experiencing very poor 
performance for file creates and the like. Throughput is not too bad, 
up to 3MB/s which is well below wire speed but still tolerable. However 
as an example a 66KB archive containing 10,000 empty files takes up to 
10 minutes to unpack; a fresh kdevelop project can take 3-4 times as 
long for a ./configure in AFS space as it does in /tmp on the same 
machine, etc.

Both clients and servers are running Debian Linux. Servers are woody, 
clients are mostly sarge.

I discovered that volumes on a non-RAIDed ext2 vice partition on our 
500MHz fileserver box were up to 3 times faster than ext3 on an 
Megaraid'ed RAID1 volume on the main 1.3GHz box. Moving the vicepa 
partitions to ext2 on the main box results in a very noticeable 
performance improvement but only from 'really dreadful' to 'dreadful'.

This might be manageable in isolation, but when many users are logged 
in the problem gets a whole lot worse. This didn't happen last year, 
and the major changes I'm aware of are the fileserver boxes have been 
upgraded from a 2.4.19 kernel to 2.4.25, and the clients have been 
upgraded from 2.4.21 to 2.4.25, and OpenAFS upgraded on the clients to 
1.2.11 (servers are 1.2.8 with patch for the Ubik time limit problem).

Suggestions? It seems that moving the main fileserver box to non-RAID 
would probably give an improvement but obviously this isn't something 
to contemplate lightly! I can provide more info if given suggestions 
about what info would be useful.

--
Phone: +64-9-373-7599 x84679	Room: 2.315, School of Engineering
Email: jj.walker@auckland.ac.nz	ICQ: 5632563 (or shout loudly)