Cache Ensemble Analysis for Understanding Algorithmic Memory Performance

We present an approach to studying ensembles of cache simulations which vary in cache features (such as size, associativity, block replacement policy, etc.) or qualities of the program (such as choice of algorithm, data storage layouts, etc.). Through the qualities of variation between their members, these ensembles can reflect the computational structure of programs and expose their performance characteristics, leading to better understanding of code and performance improvements that can be valuable in conserving scarce computing resources. We include several case studies looking at various cache performance scenarios, including some surprising performance bottlenecks in common programming operations, demonstrating the usefulness of our approach.