Dear Didier,
I am currently focusing on beam-transport optimization tasks. I have access to a virtual workstation with a substantial number of CPU cores, but I am wondering if this will represent any advantage compared to a regular laptop with 8 cores.
According to a previous post (viewtopic.php?f=3&t=819), I understand that a speed-up in computation time should be observable only for tracking but not for optimization problems. Indeed, after performing some tests in my virtual workstation, I can see that TraceWin only uses roughly 4 cores during an optimization.
Is there any way to take advantage of a higher number of available cores in optimization tasks with the current versions of TraceWin - other than, e.g., using a separate code to launch/manage ensembles of TraceWin simulations in parallel?
Many thanks in advance for your feedback and best regards,
Pau González
Parallelization capability for beam-transport optimization [SOLVED]
Re: Parallelization capability for beam-transport optimization [SOLVED]
Dear Pau,
Typically, during a tracking optimization, TraceWin will by default use 1000 particles and with this number going beyond 4 cores would be counter-productive in terms of efficiency (due to openMP). Now TraceWin can perfectly well be used in bath mode and there if your own optimisation algorithms allow parallelism and then it's up to you to see whether running 80 simulations in parallel is worthwhile.
Regards,
Didier
Typically, during a tracking optimization, TraceWin will by default use 1000 particles and with this number going beyond 4 cores would be counter-productive in terms of efficiency (due to openMP). Now TraceWin can perfectly well be used in bath mode and there if your own optimisation algorithms allow parallelism and then it's up to you to see whether running 80 simulations in parallel is worthwhile.
Regards,
Didier
Re: Parallelization capability for beam-transport optimization
Dear Didier,
many thanks for your quick feedback, I appreciate that! I agree with you that running 80 simulations in parallel would be absolutely nonsense, but I guess that maybe around 4 or 5 simulations in parallel could be convenient with the appropriate task manager. This would not speed up the computation time dramatically, of course, but a factor four would be welcome (plus would allow for more flexibility to choose the optimization algorithm, which might in turn reduce the optimization time for problems with a large number of variables...).
Best regards,
Pau
many thanks for your quick feedback, I appreciate that! I agree with you that running 80 simulations in parallel would be absolutely nonsense, but I guess that maybe around 4 or 5 simulations in parallel could be convenient with the appropriate task manager. This would not speed up the computation time dramatically, of course, but a factor four would be welcome (plus would allow for more flexibility to choose the optimization algorithm, which might in turn reduce the optimization time for problems with a large number of variables...).
Best regards,
Pau
Re: Parallelization capability for beam-transport optimization
Dear Pau
It is clear that the parallelization implemented in TraceWin was not at all designed to speed up the optimization phase, but to reduce the calculation time for very high numbers of particles, greater than 10e7.
Regards,
Didier
It is clear that the parallelization implemented in TraceWin was not at all designed to speed up the optimization phase, but to reduce the calculation time for very high numbers of particles, greater than 10e7.
Regards,
Didier