Simulation and performance evaluation of parallel software on multiprocessor systems

The paper describes a technique to simulate the execution of parallel software on a generic multiple-processor architecture and to evaluate its actual performance. Parallel programs can be written easily in C, compiled with a standard compiler and linked with a simulator library to be run on a simulated parallel system. The parallel system and its kernel primitives are simulated by the library functions, and the executable code is run in a monoprogrammed environment. (Currently both PCs and Unix machines have been used.) Communication between processes uses a message exchange synchronization mechanism similar to that of the Inmos Transputer; the use of shared memory is also allowed. Architectural parameters of the simulated system, such as the number of processors, the allocation of processes to processors, the connection scheme and the speed of operation, are software configurable. A trace utility is also programmable to show the execution of some of the kernel primitives; this feature is very important for debugging and teaching purposes. A variety of statistical parameters are collected during the execution of a program to demonstrate its behaviour on the simulated system