On Wed, 2010-11-03 at 17:36 -0400, Igor Peshansky wrote: [ . . . ] > You cannot run multi-place code with runx10 -- that script only supports > launching one place at a time.
OK, thanks -- I hadn't realized that. I should have read INSTALL.txt it is most clear on the topic! > With 2.1.0 using the pgas_sockets transport, you can either use mpiexec > from an existing MPI installation to launch multiple processes, or use > the manager/launcher combination (the process is described in the > INSTALL.txt file in the distribution). Anecdotal evidence of personal experience only, but it has to be said Chapel is way simpler to get small embarrassingly parallel codes working in parallel than X10 -- compile an executable and it executes in parallel if it can. In my case twin-Xeon workstation so eight cores. Also Chapel (as with C++ with pthreads, C++ with MPI, C++0x with futures or asynchronous function call) only uses the cores that it is told to not all the ones available. Having said that, I need to be constructive, even though this might seem adverse criticism -- which I guess it is, but be assured it is presented in an attempt to be helpful and/or find out what I am doing wrong. I tried X10 with the manager/launcher combination and as an MPI executable. Whilst launcher is happy to be called as launcher with it in the path, manager is not. manager insists on being called with an absolute path which is a serious irritant. Also I got some weird results, so I switched to treating my executable as an MPI executable -- which gives a fairer comparison to other versions I have anyway. However that gave somewhat surprising results, along with all cores getting used 100% even when not required. I appreciate that my "application" (calculating Pi by quadrature) is trivial and a microbenchmark, but it is generally a good example for presentations as it is small of code, easy to compile and run and -- most importantly -- gives something of a handle on scaling. Knowing that the X10 team has focused on the C++ back end rather than the Java back end, I am ignoring the JVM-based execution for now. (Also JIT warm up is a serious issue for JVM-based microbenchmarks at the best of times, so it is difficult to get a fair comparison for this code on that platform anyway -- though Scala performs very well.) |> runx10 Pi_X10_Parallel ==== X10 Parallel pi = 3.141592651589971 ==== X10 Parallel iteration count = 1000000000 ==== X10 Parallel elapse = 8.597293312 ==== X10 Parallel task count = 1 ==== X10 Parallel pi = 3.141592648389901 ==== X10 Parallel iteration count = 1000000000 ==== X10 Parallel elapse = 8.597158734000001 ==== X10 Parallel task count = 2 ==== X10 Parallel pi = 3.141592629477861 ==== X10 Parallel iteration count = 1000000000 ==== X10 Parallel elapse = 8.599376457 ==== X10 Parallel task count = 8 ==== X10 Parallel pi = 3.141592554064001 ==== X10 Parallel iteration count = 1000000000 ==== X10 Parallel elapse = 8.600300205 ==== X10 Parallel task count = 32 |> mpirun -n 8 Pi_X10_Parallel ==== X10 Parallel pi = 3.141592651589971 ==== X10 Parallel iteration count = 1000000000 ==== X10 Parallel elapse = 18.120695501 ==== X10 Parallel task count = 1 ==== X10 Parallel pi = 3.141592648389901 ==== X10 Parallel iteration count = 1000000000 ==== X10 Parallel elapse = 8.659647902 ==== X10 Parallel task count = 2 ==== X10 Parallel pi = 3.141592629477861 ==== X10 Parallel iteration count = 1000000000 ==== X10 Parallel elapse = 4.364804941 ==== X10 Parallel task count = 8 ==== X10 Parallel pi = 3.141592554064 ==== X10 Parallel iteration count = 1000000000 ==== X10 Parallel elapse = 2.225576907 ==== X10 Parallel task count = 32 The runx10 version is running comparable speed to all the other C++ and Chapel sequential and single core parallel versions. However the X10 MPI executed version seems to be twice as slow as all the other C++ and Chapel parallel version using either threads or MPI. So I guess my question is: why is X10 half the speed of C++/MPI? Also: does it really need to use all the cycles for the infrastructure when C++/MPI does not? If the answers to these questions are RTFM type one, feel free to just point me at the FM :-) Thanks. -- Russel. ============================================================================= Dr Russel Winder t: +44 20 7585 2200 voip: sip:russel.win...@ekiga.net 41 Buckmaster Road m: +44 7770 465 077 xmpp: rus...@russel.org.uk London SW11 1EN, UK w: www.russel.org.uk skype: russel_winder ------------------------------------------------------------------------------ The Next 800 Companies to Lead America's Growth: New Video Whitepaper David G. Thomson, author of the best-selling book "Blueprint to a Billion" shares his insights and actions to help propel your business during the next growth cycle. Listen Now! http://p.sf.net/sfu/SAP-dev2dev _______________________________________________ X10-users mailing list X10-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/x10-users
