From peterp@oak.njit.edu Thu Aug 25 15:16:20 2005 Date: Tue, 28 Jun 2005 13:37:32 -0400 (EDT) From: peter g petropoulos math fac/staff To: faculty@m.njit.edu Subject: compiler/libraries on hydra. Hello all, An update on compilers/libraries/etc. the cluster hydra.njit.edu is open for use. There are no outstanding issues that prevent users from employing it in their work. I have done a bunch of standard benchmarks, I just need to get around to making a website of the results. Just to give you an idea, the single-processor performance is in the 0.4 to 2.5 GFlop range depending on the application; the benchmarks are actually CFD codes typically used by NASA scientists. There are two queues, named "serial" (40 processors are reserved for it) and "parallel" (92 processors are reserved for it). Ignore any other queues you see reported in xpbs, they are for testing by the systems people. So, for now, the "feeder" queue submits your job to the "parallel" queue. For both serial and parallel jobs you need to use the scripts I e-mailed in early June. Somewhere on the web I saw something related to MATLAB-MPI, I would like to hear from anyone who investigates this and finds out whether Mathworks carries an MPI version of Matlab. A few words about compilers: Gnu compilers are installed (g77, gcc, g++), as well as the Portland Group compiler suite (Fortran 77/90, HPF, and C) version 5.2. This version of the compiler was used by the vendor to compile MPICH and other things related to the cluster so it will be our default compiler until we can figure out how to recompile the software that enables the cluster to operate with the newest version of the PG compilers. For those who use MPI the commands mpirun, mpiexec, mpif77, mpif90, mpicc employ v. 5.2 and set environment variables appropriately. The 64-bit PG compilers are the default (in /usr/pgi/linux86-64/5.2 ). If you need the 32-bit compilers they are in /usr/pgi/linux86/5.2/bin. Look into these directories for info, docs, examples, etc. Also, see http://www.pgroup.com/support/index.htm for further info regarding computing with the PG compilers. NOTE: environment flags are set when you use, e.g., pgf77...you may need to delve deeper into the documentation to figure out whether you need to make changes. These compilers can also do auto-parallelism. We have recently installed the newest versions of the Portland Group compilers. As it stands now they cannot be used for parallel computation but serial users can employ them by doing set path=(/usr/pgi/linux86-64/6.0/bin $path) so that the new pgf77, etc., are found first. Finally, 3 versions of the ACML library have been installed in /export/local. All versions work fine with the gnu compilers. Only version 2.5.1 in /export/local/acml2.5.1 works with the default Portland Group compilers while versions 2.5.3 and 2.6.0 (in appropriate subdirectories of /export/local) work with version 6 of the Portland compilers. ACML also includes the ScaLAPACK library for users of MPI. In the subdirectories where these libraries live there exist Release Notes and examples, pdf manuals can be found on the webpage I included in my early June e-mail. In /usr/local/vni/CTT6.0 you find the (expensive!) serial/parallel IMSL Library (instructions, examples, docs, etc.) Best regards, Peter ___________________________________________ Department: Mathematics, NJIT Office : Cullimore Hall, Rm. 608 Phone : (973)596-5626 FAX : (973)596-6467 E-mail : peterp@ouzo.njit.edu URL : http://web.njit.edu/~peterp/