Optimizing EDD, and other ineffective methods for designing heterogeneous multicores

Computer systems are inevitably defined by and limited by the metrics used to optimize them. In this talk, I will share my experience with single-ISA heterogeneous design space exploration, specifically the problem of selecting some cores for a heterogeneous processor from the set of all possible cores. Based on data from 3000 different simulated processor cores and over seven compute-years of simulation with gem5 and McPAT, I will argue that optimizing for EDD and other similar design methodologies implicitly make unrealistic assumptions about workloads. I will conclude with some ideas for better design metrics from my ongoing work.