[OpenAFS-devel] Re: [OpenAFS] openAFS on SuSE 7.2

[Harald Barth]

This is more about memcache in general.

I made the following observations for some version of Transarc 3.6
under Solaris 5.7 (the code has probably not changed that much):

*) There are complains about disk cache not existing or not enough disk
   space or even creating cache nodes on disk in spite of _mem_cache
   not using any disk blocks.

*) If my memcache was > 240Mb (I started with 512Mb on a 1000E with
   8CPUs and 1Gb of mem) the kernel panicked. This might be related that=

   the boards in a 1000E are 256Mb.

*) If my memcache was less, there were no panicks any more but I could
   see hangs once a week or so (tiresome).

Now I run the box with 300Mb of diskcache but if someone gives me some
idea what parameters should work for memcache, I'd give it another try
with 1.1.1 because with the current setup I get hangs about once every
few month. The box sees quite a lot of traffic (ftp.stacken.kth.se,
www.stacken.kth.se, ftp.se.kernel.org). Or I might give it a try
with Arla first. Hm.

> I seem to recall some magic with the mem-cache and sizes.

Ugh. I want to talk to the magician on duty :-)

J=F6rg:

I do not think that there are enough people around running memcache so
that it could be considered well tested. Stange idea: Make a memory
file system with ext2 in it and then run a "disk cache" against that
if you need the performance. However I have not done that myself.

Harald.