[Prev][Next][Index][Thread]

Re: caching - which machine should we use?



On Thu, 6 Jul 1995 rwald@dwst12.edvz.sbg.ac.at wrote:
>   
>   does someone know how many performance one process of the cern httdp needs?
>   has anyone expirience which machine we should use (buy) or should
>   we use 2-4 server?

For interest, we are running a SPARC 5 32MB Ram, 8.4GB cache and CERN 3.0
We peak at just under 100,000 accesses per day - I think about 9000 
requests per hour is the most we've seen.

This machine is very busy - iostat regularly shows the caching disk 90% 
busy.  It provides a "national" cache for a few universities and CSIRO in 
Australia.  You can see usage stats at http://www.unimelb.edu.au/info-stats/
(thanks to Martin Gleeson at the University of Melbourne).

I dont think this machine could service a bigger load - we'd hoped to use 
Harvest instead of CERN by now, but there are a few problems (Danny has 
mentioned them on this list recently I think).

With this size cache, we are getting good hit rates (hovering between 
30-35%), but most of the code using "long integers" on the Sparc cant 
cope well with cache sizes that take more than 32 bits to represent!  
Fortunately, this seems to only effect reporting, not normal operation.

Kent Fitch                           Ph: +61 6 276 6711
ITSB   CSIRO  Canberra  Australia    kent.fitch@its.csiro.au
"There is nothing like the certainty of a bourgeois income to promote
 bohemian behaviour"  - George Bernard Shaw




References: