Hi, On Tue, Mar 15, 2016 at 08:47:52AM +0000, Alf Birger Rustad wrote: > Thanks for the announcement! There is one issue that I hope > everybody with informed opinions can contribute to. Today, writing > results is a performance bottleneck flow. Hence, we should find a > way for flow to write results asynchronously, i.e., the simulator > should not need to wait computing the next timestep while the > results are written. It has been suggested to split out the writing > in a separate process to accomplish this, but I am not sure how this > could be done in a good way. Implementing the writer as a server > seems like over-engineering, and we probably do not want to mandate > MPI to run the simulator. The simplest option I can think of is > splitting the writing to a new thread with openmp.
As I am the one who brought this up, I have to relativate this a bit. I experienced this bottleneck for a parallel run with 8 cores. The time spent for output was 31% of the total time. For a sequential run it is probably below 10% but I have not measured it. Minimizing the time for the ouput of a sequential run as proposed comes at cost, too. We need to duplicate the memory. What I am concerned about is that the output is not parallel at all which will bite us more when using more cores. Having a dedicated output core is just a first step. Ultimately, we need several cores doing output in parallel in a parallel file format. But this can wait some time, I guess. Cheers, Markus -- Dr. Markus Blatt - HPC-Simulation-Software & Services http://www.dr-blatt.de Hans-Bunte-Str. 8-10, 69123 Heidelberg, Germany, USt-Id: DE279960836 Tel.: +49 (0) 160 97590858
signature.asc
Description: Digital signature
_______________________________________________ Opm mailing list Opm@opm-project.org http://opm-project.org/cgi-bin/mailman/listinfo/opm