Slices: Provisioning Heterogeneous HPC Systems

Abstract

High-end computing systems are becoming increasingly heterogeneous, with nodes comprised of multiple CPUs and accelerators, like GPGPUs, and with potential additional heterogeneity in memory configurations and network connectivities. Further, as we move to exascale systems, the view of their future use is one in which simulations co-run with online analytics or visualization methods, or where a high fidelity simulation may co-run with lower order methods and/or with programs performing uncertainty quantification. To explore and understand the challenges when multiple applications are mapped to heterogeneous machine resources, our research has developed methods that make it easy to construct 'virtual hardware platforms' comprised of sets of CPUs and GPGPUs custom-configured for applications when and as required. Specifically, the 'slicing' runtime presented in this paper manages for each application a set of resources, and at any one time, multiple such slices operate on shared underlying hardware. This paper describes the slicing abstraction and its ability to configure cluster hardware resources. It experiments with application scale-out, focusing on their computationally intensive GPGPU-based computations, and it evaluates cluster-level resource sharing across multiple slices on the Keeneland machine, an XSEDE resource.