Learning not to share

摘要

Strong reasons exist for executing a large-scale discrete-event simulation on a cluster of processor nodes (each of which may be a shared-memory multiprocessor or a uniprocessor). This is the architecture of the largest scale parallel machines, and so the largest simulation problems can only be solved this way. It is a common architecture even in less esoteric settings, and is suitable for memory-bound simulations. This paper describes our approach to porting the SSF simulation kernel to this architecture, using the Message Passing Interface (MPI) system. The notable feature of this transformation is to support an efficient two-level synchronization and communication scheme that addresses cost discrepancies between shared-memory and distributed memory. In the initial implementation, we use a globally synchronous approach between distributed-memory nodes, and an asynchronous shared-memory approach within a SMP cluster. The SSF API reflects inherently shared-memory assumptions; we report therefore on our approach for porting an SSF kernel to a cluster of SMP nodes. Experimental results on two architectures are described, for a model of TCP/IP traffic flows over a hierarchical network. The performance on a distributed network of commodity SMPs connected through ethernet is seen to frequently exceed performance on a Sun shared-memory multiprocessor.