[OpenAFS-devel] good and bad performance of memory cache

[Quanah Gibson-Mount]

--On Wednesday, February 05, 2003 12:43 PM -0800 Edward Moy 
<emoy@apple.com> wrote:

> Since I was starting to investigate memory caching, I decided to run
> iozone to really exercise the cache.  With disk caching on my  moderately
> slow (400MHz) test machine, the full iozone test suite took  nearly seven
> hours, with the largest files (going up to half a  gigabyte) really
> dragging down performance.  So much so that iozone  running over NFS2
> actually runs 20 minutes faster than AFS.
>
> I ran the same test with memory caching enabled (and even a larger
> cache; 80 MB memory cache versus the 30 MB disk cache).  The whole test
> suite now took nearly *twelve* hours.
>
> In comparing results, speed is definitely faster for the memory cache
> when the file size is smaller than the cache size.  But when the file
> size is larger than the memory cache, performance is often worse (in
> some cases more that twice as bad) as the disk cache.
>
> Anyone have any ideas about why this is so?

Edward,

I'll add a comment on our own experience of using the AFS memory cache. 
The memory cache has a tendency to keep stale data rather than update 
changes when a file is accessed.  The disk cache is much better about 
refreshing things that have changed.  We noticed this with our webserver, 
where people's pages (served out of AFS) would get stale daily, and it 
would require a reboot of the server to have the data refreshed.

--Quanah

--
Quanah Gibson-Mount
Senior Systems Administrator
ITSS/TSS/Computing Systems
Stanford University
GnuPG Public Key: http://www.stanford.edu/~quanah/pgp.html