From: Marc L. Smith (mlsmith_at_colby_dot_edu)
Date: Mon Aug 15 2005 - 12:50:07 PDT
On Aug 15, 2005, at 3:21 PM, jcduell_at_lbl_dot_gov wrote: > Thanks for the feedback, Marc. > > I notice that UDP is the default network for your install. We've > improved our documentation for how to run UDP-based programs in this > beta, and UPC executables now provide a helpful message if you try to > run them over UDP without the right setup (i.e., without telling the > runtime which hosts to run the job on). I haven't noticed any complaints at runtime, but just to be sure... I set my environment vars sets as follows: GASNET_SPAWNFN=S SSH_SERVERS="n1 n2 n3 n4 n1 n2 n3 n4" I listed n1-n4 in twice, because each of my four nodes has two processors. I ordered them as I did because it appeared to determine allocation of threads to processors, and I wanted even fan-out (e.g., if I ran a program with 4 threads, and listed servers via "n1 n1 n2 n2...", then the four threads were allocated by default, 2 per node, to nodes n1 and n2. Is this correct? > I remember you were thinking of using UPC for a course you'll be > teaching. We'd love to hear how that goes. Please have have your > students send us bug reports for any issues they encounter (extra > credit > for each bug reported?) Yes, in fact I've chosen the UPC text by El-Ghazawi. The course won't be purely UPC programming, but UPC is one of the main languages students will learn to program in. I've been through the UPC text and think it is, like UPC itself, and K&R, a very concise, useful, and elegant book. I'm sure GCC's support of UPC will encourage students that UPC is not merely an academic language -- though I'm going to take your advice and stick with Berkeley for translation and runtime. :-) I'll also be using TCP Linda (by Scientific Computing Associates) as another example of a distributed shared memory model. I think together, UPC and Linda/Tuple Space models will give students a broad sense of shared memory concurrency. For message passing, we'll be using either JCSP or occam (two realizations of Hoare's CSP model of concurrency). I've opted not to cover MPI. I hope I don't offend anyone on this list, but I fear MPI is so large it is overwhelming for students. In contrast, UPC and Linda are quite elegant, as are JCSP and occam, with only a few programming constructs to learn and master. Thanks for all your support. I'm not anticipating any problems with UPC, but I'll let you know how things are going, and be sure to pass along any perceived bugs we encounter. Cheers, Marc > Cheers, > > -- > Jason Duell Future Technologies Group > <jcduell_at_lbl_dot_gov> Computational Research Division > Tel: +1-510-495-2354 Lawrence Berkeley National Laboratory > >