Constant-work multiprogram throughput metrics for microarchitecture studies

Multicore processors can improve performance by decreasing the execution latency of parallel programs, or by increasing throughput, i.e., the quantity of work done per unit of time when executing independent tasks. Throughput is not necessarily proportional to the number of cores and can be impacted significantly by resource sharing in several parts of the microarchitecture. Quantifying the impact of resource sharing on throughput requires a throughput metric. A majority of microarchitecture studies use equal-time throughput metrics, such as IPC throughput or weighted speedup, that are based on the implicit assumption that all the jobs execute for a fixed and equal time. We argue that this assumption is not realistic. We propose and characterize some new throughput metrics based on the assumption that jobs execute a fixed and equal quantity of work. We show that using such equal-work throughput metric may change the conclusion of a microarchitecture study.