Criticality-driven superscalar design space exploration

Abstract

It has become increasingly difficult to perform design space exploration (DSE) of computer systems with a short turnaround time because of exploding design spaces, increasing design complexity and long-running workloads. Researchers have used classical search/optimization techniques like simulated annealing, genetic algorithms, etc., to accelerate the DSE. While these techniques are better than an exhaustive search, a substantial amount of time must still be dedicated to DSE. This is a serious bottleneck in reducing research/development time. These techniques do not perform the DSE quickly enough, primarily because they do not leverage any insight as to how the different design parameters of a computer system interact to increase or degrade performance at a design point and treat the computer system as a black-box.We propose using criticality analysis to guide the classical search/optimization techniques. We perform criticality analysis to find the design parameter which is most detrimental to the performance at a given design point. Criticality analysis at a given design point provides a localized view of the region around the design point without performing simulations at the neighboring points. On the other hand, a classical search/optimization technique has a global view of the design space and avoids getting stuck at a local maximum. We use this synergistic behavior between the criticality analysis (good locally) and the classical search/optimization techniques (good globally) to accelerate the DSE.For the DSE of superscalar processors on SPEC 2000 benchmarks, on average, criticality-driven walk achieves 3.8x speedup over random walk and criticality-driven simulated annealing achieves 2.3x speedup over simulated annealing.