[gate-users] Gate/Condor users

[Manuel Bardiès]

Dear all,

Some of us are using Condor to distribute Gate jobs on a cluster.
Have you tried to benchmark the performances ?
For example, if MyMacro.mac needs 20 hours to run, an 'ideal' 
simulation time for a Condor distributed task on 20 nodes would be 1 
hour. Obviously this is impossible, especially since all nodes may not 
be available at the start of a simulation, and there is always a 
decrease in performances when splitting a macro in several jobs (some 
time - initialisation for example- is needed on all jobs, regardless of 
the number of simulated particles).

Our (recent) experience with Condor - and the way we split calculations 
- is that one doesn't get even close to the 'perfect' case. An example 
: for a simulation involving 16 nodes in SPECT, we get the efficiency 
of 10 'ideal' machines (i.e. computing time is equal to computing time 
for a single machine divided by 10 instead of 16).
I'm not very happy with that and I wonder if somebody could provide 
other figures.
Maybe (surely) we can tune our macro splitting and job management...

All the best,

Manu

Manuel Bardiès
INSERM UMR 601
9 Quai Moncousu
44093 Nantes cedex
-----------------------------
Tel:   02 40 41 28 21
Fax:  02 40 35 66 97
Sec:  02 40 08 47 47