Performance Models for Task-based Scheduling with Disruptive Memory Technologies

Disruptive memory technologies break out of the memory pyramid and mandate specialized performance models and algorithms for optimal use. While the literature offers various - occasionally conflicting - models for individual technologies, few of these take NUMA effects into account, and there is little support for performance-aware online scheduling on systems that combine several technologies. This paper examines whether technology- and NUMA-aware performance models for such systems are feasible, using high-bandwidth memory and near-memory computing as sample technologies. We determine which run-time parameters affect compute and memory performance, build performance models, and incorporate them into the HetSim simulator for scheduling on heterogeneous hardware. We show that these models enable HetSim to simulate CPU and NMC performance with an average makespan error of 14 %. While our HBM models are also accurate within 14 %, we find that HBM placement decisions are infeasible without knowledge about a task's memory access patterns. Moreover, HBM is not a universal improvement: it can in fact slow down applications.