How to apply Amdahl's law to multithreaded multicore processors

摘要

In this paper we comment on a recent article on Amdahl's law for multithreaded multicore processors. We propose that models for predicting speedup in such systems should explicitly separate the memory and computational parts of a workload. On the one hand, not doing so can lead to speedup calculations with more than one possible result. We show this for a simple example which exploits cache affinity between two sequential tasks. On the other hand, we show that explicit separation of the workload in this case leads to a unique result. We contend that decomposition of workloads into computational and memory parts can resolve similar ambiguity in general and is fundamental since it is apparent in the Turing Machine model of computation. Further research into this separation in the formulation of Amdahl's law for modern architectures is advocated.