Hello,
I did a quick test and ran the same simulation including an optimization process with PARTRAN using two different platforms :
- Dell mobile workstation with intel core i7-11850H at 2.5GHz/RAM 14.8 Go under Fedora Linux 37 Workstation Edition -> it took 32,11 min
- Server with 48 intel Xeon Gold 5220s at 2.7GHz/RAM 678 Go under CentOS Linux release 7.9.2009 (Core) -> it took 1.99 hours (no other users on the server).
In both case, the option "Using MultiThreading capabilities" was selected.
For both calculation, Tracewin uses 4 threads (that can be seen in Help-> About).
Maybe it's a limitation due to running Tracewin under Linux ?
Emmanuel
Performance speed low under linux
Re: Performance speed low under linux
Dear Emmanuel,
The OpenMP-based parallelization is not at all complete and only deals with tracking. However, it is only effective for a high number of particles. Below 1000, the speed-up even becomes unfavourable beyond 4 cores. The optimisation process in general deliberately limits the number of particles to 1000 to speed up the simulations, which is why you observe this number of cores. If you run a 10e6 particle simulation you should see the number of cores used increase.
Regards,
Didier
The OpenMP-based parallelization is not at all complete and only deals with tracking. However, it is only effective for a high number of particles. Below 1000, the speed-up even becomes unfavourable beyond 4 cores. The optimisation process in general deliberately limits the number of particles to 1000 to speed up the simulations, which is why you observe this number of cores. If you run a 10e6 particle simulation you should see the number of cores used increase.
Regards,
Didier
Re: Performance speed low under linux
Hello Didier,
10⁴ particles were simulated. I guess that the time taken for each simulation should be at least roughly the same, if not equal. One would even expect the Xeon to produce faster calculations than the i7.
I foresee to use the server for error studies but I wonder if it's worth doing this as its 48 processors (18 cores) don't even match a single i7.
Emmanuel
10⁴ particles were simulated. I guess that the time taken for each simulation should be at least roughly the same, if not equal. One would even expect the Xeon to produce faster calculations than the i7.
I foresee to use the server for error studies but I wonder if it's worth doing this as its 48 processors (18 cores) don't even match a single i7.
Emmanuel
Re: Performance speed low under linux
Dear Emmanuel,
In error studies, there is no more parallelization. Each run uses 100% of each core. If you have 48 cores, you should declare 48 cores in the TraceWin interface to be as efficient as possible.
Now to really compare your 2 machines, you should uncheck the parallelisation option and see for a single run, if there is such a difference between the i7 and the Xeon. I don't see such a gap on my side, but beware here the frequency obviously counts a lot.
Regards,
Didier
In error studies, there is no more parallelization. Each run uses 100% of each core. If you have 48 cores, you should declare 48 cores in the TraceWin interface to be as efficient as possible.
Now to really compare your 2 machines, you should uncheck the parallelisation option and see for a single run, if there is such a difference between the i7 and the Xeon. I don't see such a gap on my side, but beware here the frequency obviously counts a lot.
Regards,
Didier