The virtual-time data-parallel machine

Abstract

The authors propose the virtual-time data-parallel machine to execute SIMD (single instruction multiple data) programs asynchronously. They first illustrate how asynchronous execution is more efficient than synchronous execution. For a simple model, they show that asynchronous execution outperforms synchronous execution roughly by a factor of (ln N), where N is the number of processors in the system. They then explore how to execute SIMD programs asynchronously without violating the SIMD semantics. They design a first in, first out (FIFO) priority cache, one for each processing element, to record the recent history of all variables. The cache, which is stacked between the processor and the memory, supports asynchronous execution in hardware efficiently and preserves the SIMD semantics of the software transparently. Analysis and simulation results indicate that the virtual-time data-parallel machine can achieve linear speed-up for computation-intensive data-parallel programs when the number of processors is large.<>