Exploring Mapping Strategies for Co-allocated HPC Applications

Abstract

In modern HPC systems with deep hierarchical architectures, large-scale applications often struggle to efficiently utilize the abundant cores due to the saturation of resources such as memory. Co-allocating multiple applications to share compute nodes can mitigate these issues and increase system throughput. However, co-allocation may harm the performance of individual applications due to resource contention. Past research suggests that topology-aware mappings can improve the performance of parallel applications that do not share resources. In this work, we implement application-oblivious, topology-aware process-to-core mappings via different core enumerations that support the co-allocation of parallel applications. We show that these mappings have a significant impact on the available memory bandwidth. We explore how these process-to-core mappings can affect the individual application duration as well as the makespan of job schedules when they are combined with co-allocation. Our main objective is to assess whether co-allocation with a topology-aware mapping can be a viable alternative to the exclusive node allocation policies that are currently common in HPC clusters.