Date: February 8, 1999
Creator: Benger, W.; Foster, I.; Novotny, J.; Seidel, E.; Shalf, J.; Smith, W. et al.
Description: We have found that the hardware and software infrastructure exists to simulate general relativity problems in a distributed computational environment, at some cost in performance. We examine two different issues for running the Cactus code in such a distributed environment The first issue is running a Cactus simulation on multiple parallel computer systems. Our objective is to perform larger simulations than are currently possible on a single parallel computer. We distribute Cactus simulations across multiple supercomputers using the mechanisms provided by the Globus toolkit. In particular, we use Globus mechanisms for authentication, access to remote computer systems, file transfer, and communication. The Cactus code uses MPI for communication and makes use of an MPI implementation layered atop Globus communication mechanisms. These communication mechanisms allow a MPI application to be executed on distributed resources. We find that without performing any code optimizations, our simulations ran 48% to 100% slower when using an Origin at the National Center for Supercomputing Applications (NCSA) and an Onyx2 at Argonne National Laboratory (ANL). We also ran simulations between Cray T3Es in Germany and a T3E at the San Diego Supercomputing Center (SDSC). Running between the T3Es in Germany resulted in an increase in execution time ...
Contributing Partner: UNT Libraries Government Documents Department