[OpenAFS] AFS client suddenly extremely slow

Jeffrey Hutzelman jhutz@cmu.edu
Wed, 01 Nov 2006 13:47:10 -0500

On Tuesday, October 31, 2006 05:20:05 PM -0500 Jim Rees <rees@umich.edu> 

>   So just be be 100% clear, to turn autotune on, all you need to do is
>   start up "afsd" without any of these options: "-stat, -dcache,
>   -daemons, -volumes, -chucksize"?
> That's right.  You can find out what the values got auto-tuned to with
> "cmdebug -cache".

Note that it's not an all-or-nothing deal, and in fact, it has always been 
the case that afsd would compute default values for whichever parameters 
you didn't provide.  All that changed recently was that the way those 
values are calculated was significantly improved.  As the person who came 
up with most of the current formulas, I believe they're fairly good, but 
there is room for improvement.  Suggestions are welcome.

>   Does auto-tuning work only for disk cache, or also with "-memcache"?
> I think it's disk cache only.

When you are using an in-memory cache, there is a fixed relationship 
between the cache size, chunk size, number of chunks, and number of dcache 
entries.  You must provide the cache size, and can optionally set either 
the chunk size or number of dcache entries, but not both.  If you don't 
give either one, then afsd will use a default chunk size of 8K, so as to 
try to make efficient use of memory.

The default values for the number of vcache and volume cache entries are 
computed in the same way for both memory and disk caches.  Note that for a 
memory cache, this will result in a fairly small value; this is probably a 
case where we should make some improvements to insure you don't end up with 
too few vcache entries, even if the cache is very small.

-- Jeff