Predicting Multiprocessor Memory Access Patterns with Learning Models

摘要: Machine learning techniques are applicable to computer system optimization. We show that shared memory multiprocessors can successfully utilize machine algorithms for access pattern prediction. In particular three different on-line prediction were tested learn and predict repetitive patterns typical parallel processing applications, the 2-D relaxation algorithm, matrix multiply Fast Fourier Transform on a multiprocessor. The predictions then used by routing control algorithm reduce latency in interconnection network configuring provide needed paths before they requested. Three trainable tested: 1). Markov predictor, 2). linear predictor 3). time delay neural (TDNN) predictor. Different predictors performed best but TDNN produced uniformly good results.