A Pluggable Framework for Composable HPC Scheduling Libraries

摘要

Driven by the increasing diversity of current and future HPC hardware and software platforms, the HPC community has seen a dramatic increase in research and development efforts into the composability of discrete software systems. While modularity is often desirable from a software engineering, quality assurance, and maintainability perspective, the barriers between software components often hide optimization opportunities. Recent examples of work in composable HPC software include GPU-Aware MPI, OpenMP's target directive. Lithe, HCMPI, and MVAPICH's unified communication runtime. These projects all deal with breaking down the walls between software or hardware components in order to achieve performance, programmability, and/or portability gains. However, they also generally focus on composing only specific types of HPC software and have limited extensability. In this paper, we present work on using a pluggable API framework on top of a "generalized work-stealing" runtime to achieve composability of communication, accelerator, and other HPC libraries. We motivate this work by the increasing heterogeneity of HPC hardware, software, and applications, and note that as heterogeneity increases many discrete software frameworks will need to cooperate within a single process. Our framework, called HiPER (a Highly Pluggable, Extensible, and Re-configurable scheduling framework for HPC) enables exactly this cooperation. We demonstrate the programmability improvements enabled by the HiPER framework through the use of novel APIs which reduce programmer burden. We also present performance studies that demonstrate that through unified and asynchronous scheduling of composed software systems we can achieve performance improvements over hand-optimized benchmarks.