sow@cad.luth.se (Sven-Ove Westberg) (04/11/91)
I have no experiance at all from SGI so please forgive me if this is a too simple question for this group. I am very interested to know how SGI utilize the Multicpu architecture on a computational intencive task eg FEA. Can the job be split over the cpus? How is it done? Did the compiler do it automaticly or did I have to do it my self? What is the resulting performace? Did anyone have experience with using multicpus on one job? Sven-Ove Westberg, CAD, University of Lulea, S-951 87 Lulea, Sweden.
mas@everest.corp.sgi.com (Michael A. Schulman) (04/11/91)
In article <1784@eru.mt.luth.se>, sow@cad.luth.se (Sven-Ove Westberg) writes: |> |> I have no experiance at all from SGI so please forgive me if this |> is a too simple question for this group. |> |> I am very interested to know how SGI utilize the Multicpu architecture |> on a computational intencive task eg FEA. Can the job be split over |> the cpus? How is it done? Did the compiler do it automaticly or did |> I have to do it my self? What is the resulting performace? |> Did anyone have experience with using multicpus on one job? |> |> Sven-Ove Westberg, CAD, University of Lulea, S-951 87 Lulea, Sweden. Yes, the job can be split over multiple CPU's. It can be done in a variey of ways, either using a FORTRAN pre-processor to determine which loops can be split, or by inserting these drectives by yourself. In addition, there are different levels of libraries that the programmer may use to gain increased control over the parallelism. Performance on a multiple cpu system may vary, and some linear speedups have been noted. On one commercial FEA program, we have acheived a 3.2X speedup on a 4 processor system, start to finish, not just the parallel portion. -- Michael Schulman Silicon Graphics Computer Systems 2011 N. Shoreline Blvd. Mountain View, CA 94039 mas@sgi.com