>
> The time required to reclaim (recycle) space in the cache increases
> linearly with the number of chunks. A memory-based cache does not
> usually improve performance; it was implemented solely to support
> diskless workstations.
>
We were having similiar problems as the original poster's on one of our
large interactive systems. Our basic probelm was that the system would
come to alomst a complete halt when we 5 or so users executing
links. We had the cache size set up to 1G and we also played with chunk
sizes and number of files. Finally since we had enough memory to do so, we
increased the amount of cache kept in memory. This instantly improved the
situtation. The current config we are running with is:
GRANDE="-stat 2800 -dcache 4608 -daemons 7 -volumes 256"
>From my notes I have down that dcache indicates the number of cache item
entries you want cached in memory. We set this to 75% of the # of vfiles
in the cache. There is a formula we used to figure out the # of vfiles
in the cache
((<cachesize> * 1024)/ <chunksize>) * 1.5 = # vfiles in cache
Where we set:
cachesize = 256
chunksize = 64
This formula MAY be in the AFS sys admin guide but I can't find it now.
I know we did reference some things out of the guide when setting this
up. In looking at the guide now I found that chapter 13 had some better
definitions of the terms I used above. You might want to look at it.
Note that we found it very difficult to get assistance from Transarc on
the tuning of the parameters above and finally had a Transarc consultant
come to our site for a two day consulting stint ($$). He helped us tune the
cache parameters to what we have above, in about a half hour. This is just a
strong pitch for more technical assistance on issues like this via phone
support.
We have not played with the memory cache very much but a few folks
are using it on desk top systems and they said performance for them
improved. We have not tried on any of our large multi-user MP machines.
Lisa Giacchetti
FERMI National Accelerator Lab
[EMAIL PROTECTED]
> I'm pretty sure that you can "spiff up response" through appropriate
> tuning, but acheiving optimal performance requires some study and
> analysis.
>
> Are you presently tuning any of the afsd switches? Using "rc.large",
> for instance? If not, then I can definitely guarantee some
> improvements. If so, try also adding "-files 2800".
>
> Do your users log out and log in frequently? If so, consider adding
> "unlog" to their logout procedure.
>
