[OSGeo-Discuss] WMS Performance Shootout presentation/results

Andrea Aime aaime at opengeo.org
Fri Oct 23 19:50:04 PDT 2009

Craig Miller ha scritto:
> I agree wholeheartedly.  It looks like the bottleneck was the database.  
> I’ve been privy to some MapServer tests done by testing teams over 
> several months and the result there was always deploying the data with 
> long update cycles to the middle tier disks instead of using the 
> database.  Only then could the performance of the actual map servers be 
> evaluated.  Performance shootouts/testing take time to do correctly as 
> each run teaches you more and more about how your deployment 
> architecture affects the results.

Actually running the benchmarks it looked the CPU was the bottleneck,
we had 25% cpu load in the single test (so 100% of the one CPU
used by that test, out of the 4 available) and during the higher loads
the CPU usage was jumping between 80% and 100% (so almost all 4 cpu
maxed out).

The database server was hardly using more than 25-30% of its CPU
and the network shouldn't have played the bottleneck either (past
experience suggests it takes a tilecache to actually turn a 100Mbit
line into a bottlenek), thought we don't have numbers for that

During the benchmark I had very little time to run profilers, but
the few times I've tried in GeoServer the time seemed to be splitted
quite equally between data fetching, actual drawing, and output
image encoding... which is kind of the worst thing you can get out
of a profile run, since it does not point to any culprit.

Getting a profile out of MapServer is surely going to be interesting,
the codebase is smaller and there is no garbage collection going on
so the results should be clearer. Looking forward to see those results :-)


More information about the Discuss mailing list