Near fine grain parallel processing of circuit simulation using direct method

This paper proposes a parallel processing scheme of electronic circuit simulation using the direct method and evaluates the performance of the scheme on a multiprocessor system named OSCAR (Optimally SCheduled Advanced MultiprocessoR) with centralized shared memory, distributed shared memory and local memory. A special purpose compiler is used to realize efficient near fine grain parallel processing of circuit simulation on an actual multiprocessor system. The compiler automatically generates a circuit simulation program from SPICE format input data. The compiler decomposes the program into tasks, analyzes data dependencies among the tasks, schedules the tasks and generates optimized parallel machine code. In the parallel machine code generation process, a different machine code for each processor, which consists of task codes, data transfer codes and synchronization codes, is generated by using static scheduling results.