Spatial instruction scheduling for raw machines

Abstract

Instruction scheduling on software exposed architectures, such as Raw, must be performed in both time and space. The complexity and variance of application scheduling regions dictates that the space-time scheduling task be divided into phases. Unfortunately, the interaction of phases presents a phase ordering problem. In this thesis, the structure of program scheduling regions is studied. The scheduling regions are shown to have varying characteristics that are too diverse for a single simple algorithm to cover. A new scheduling technique is proposed to cope with this diversity and minimize the phase ordering problem. First, rather than maintaining exact mappings of instructions to time and space, the internal state of the scheduler maintains probabilities for different assignments of instructions to time and space resources. Second, a set of small scheduling heuristics cooperatively iterate over the probabilistic assignments many times in order to minimize the effects of phase ordering. A simple spatial instruction scheduler for Raw machines based on this technique is implemented and shown to outperform existing spatial scheduling systems on average.