On Thu, Sep 04, 2008 at 04:06:31PM -0400, David Bauer wrote: > taskPR was an attempt to get 'free' parallelism out of already > existing programs by using simple data dependencies to figure out > which individual statements in a program can be run in parallel. > The name comes from the description of the program as exploiting > task-level parallelism.
Ah, and thus your reference to Tomasulo's algorithm, interesting. Thanks for straightening me out there. http://users.ece.gatech.edu/~gte810u/Parallel-R/ > (If anybody actually uses or has successfully used this package, I > would love to hear about it, btw. While the package *does* work, > there are probably few cases where it is worth it.) What would you say typically limits taskPR's approach, not finding enough instruction-level parallelism at the R script level, or the communications overhead (probably latency) of trying to make use of it? If latency, then perhaps taskPR would work better in a multi-threaded R interpreter, rather than across a TCP/IP network fabric. To roughly test that empirically (assuming you are in fact using MPI for the communications), I suppose you could start up your several R processes on a single fat SMP node, and use an MPI that sends messages through fast shared memory. That's probably still slower than thread-to-thread communications, but it should be much lower latency than TCP/IP. Maybe you already tried something like that? -- Andrew Piskorski <[EMAIL PROTECTED]> http://www.piskorski.com/ ______________________________________________ R-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
