From: Lars Rzymianowicz Newsgroups: comp.parallel.mpi Subject: Re: Performance of overlapped communication and computation Date: Fri, 08 Jan 1999 09:30:28 +0100 Organization: Dept. of Computer Engineering, University of Mannheim, Germany Message-Id: <3695C224.4557861A@mufasa.informatik.uni-mannheim.de> References: <76tkr7$efc$1@lux.doc.ic.ac.uk> <76tn46$ni3$1@pegasus.csx.cam.ac.uk> <7705o8$jf$1@lux.doc.ic.ac.uk> <770p7a$cbm$1@nnrp1.dejanews.com> <77276u$rfp$1@lux.doc.ic.ac.uk> <772v8p$8q7$1@nnrp1.dejanews.com> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit don_schad@my-dejanews.com wrote: > Perhaps you might not want to bother. Here are the results of running > the 3 versions of your code on an 4-CPU SGI Challenge RS-10000 super-dooper > computer: > > isend/irecv/wait: > [0] all done, tick = 4017432, time = 42.8943 > [1] all done, tick = 4022494, time = 42.8962 > > send/recv blocking > > [0] all done, tick = 4027718, time = 42.6865 > [1] all done, tick = 4034738, time = 42.7301 > > MPI_Sendrecv() > > [0] all done, tick = 4015940, time = 42.8822 > [1] all done, tick = 4015395, time = 42.8825 > > For some reason the blocking version seems to run faster, > but perhaps this is due to fewer function calls or something? > The difference is trivial (I repeated runs and we see the > same thing). Hi all, i think, the problem is here, that the test program is not an adequate benchmark for blocking vs non-blocking calls. What we have is an app, where all processes do the same amount of work, and all tests were run with equal-strong nodes. So all nodes will reach their MPI calls at more or less the same time... Non-blocking calls can help a lot in a more heterogenous environment (different CPU clock, RAM size ...). And if processes have different amount of work. Think of a master-slave model, where the master distributes different data portions to the slaves. You won't use blocking calls here, or? In general i would say, if you don't need synchronization, better use nb calls. The effect is even larger with HPC networks like Myrinet, ServerNet etc., where a lot of work is done by the network interface and not the CPU. Lars -- Homepage: http://mufasa.informatik.uni-mannheim.de/lsra/persons/lars/