[OpenAFS] 1.3.84 disk/memory cache status & memcache corruption

Andrej Filipcic andrej.filipcic@ijs.si
Tue, 07 Jun 2005 14:48:25 +0200


Hi,

I have made few stress tests of openafs 1.3.84 on gentoo amd64 systems,
linux 2.6.11.x. On non-smp machines, both disk and memory cache work
well. On smp machine (dual opteron 250) disk cache works, but memory
cache has some problems resulting in corruption by write.

I did the following tests:
1) parallel rsync  (10 processes) of linux source to new directories,
everything on afs
2) make -j16 of linux kernel in afs space


1) worked OK for both uni processor and smp kernels with disk or memory
cache.
2) works for disk/memory cache on uni kernels and disk cache on smp kernels.

afsd options used for memory cache are
-nosettime -dynroot -memcache -blocks 65536 -chunksize 16 -stat 2800
-daemons 3 -volumes 128

On smp kernel, the following happens with 2). After compilation of few
10 files,
there is a corrupted object file. It is corrupted on fileserver, not in
cache. With
make -j1, the same happens but quite rarely. So it seems, there is some
kind of
write inconsistency with heavy load. I have no idea how to explore that
in more
detail.

So in summary, memcache on smp does not work on heavy parallel I/O.

Cheers,
Andrej

-- 
_____________________________________________________________
   doc. dr. Andrej Filipcic,   E-mail: Andrej.Filipcic@ijs.si
   Department of Experimental High Energy Physics - F9
   Jozef Stefan Institute, Jamova 39, P.o.Box 3000
   SI-1001 Ljubljana, Slovenia
   Tel.: +386-1-477-3674    Fax: +386-1-477-3166
-------------------------------------------------------------