On Extending Parallelism to Serial Simulators

15 We are continuing to extend the capabilities of U.P.S. , and seek to expand its user base and to port it to other platforms. We are implementing an MPI (Message Passing Interface) version; MPI is the newly emerging standard for message-passing applications. When operational, U.P.S. will execute on any platform supporting both CSIM and MPI, which we expect to include most parallel systems as well as networks of workstations. In addition, we are particularly interesting in porting it to shared-memory multiprocessor workstations, as these are increasingly common. Acknowledgments We gratefully acknowledge the contribution of Cathy Roberts who developed the detailed CSIM model of a switching center, and of Phil Dickens who ported CSIM to the i860 processor. Processors 1 2 4 8 16 32 64 0 1 2 3 4 normalized procs x time Figure 7: Normalized processor time product, illustrating super-linear accelerations on large numbers of processors. to adopt parallel simulation until there is clear evidence of a need, and clear evidence of an ability to satisfy that need. We propose here a methodology for providing extensions to existing simulators, extensions that do not require modiication to those simulators. The simulation modeler uses the tool as normal, except he partitions the model into sub-models, one per processor, and incorporates the extension constructs into those sub-models. The extensions are designed to provide essentially the same functionality as basic constructs in the base simulator, but to do so in a way that the inter-processor synchronization and communication is automated. The distributed model looks very much like a non-distributed model. In U.P.S. , a small set of additional calls provide the parallel simulation capability. Almost all of the model is ordinary C and CSIM. The methodology very much involves the modeler, who must partition the model, and must consider the performance tradeoos between diierent synchronization strategies when performing that partition. However, the simulation modeler does not have to implement the communication and synchronization. That is done once, by the extension's author. This provides industry with a lower risk path to parallel processing, and an ability to experiment with it with low software cost. In view of the general perception of parallel simulation as tricky business, we believe this approach proves an important rst step towards making parallel simulation safe for the masses. We are exploring this methodology by writing an extension library, U.P.S. , for the commercial tool CSIM (now distributed …