RII Track-4: NSF: Scalable MPI with Adaptive Compression for GPU-based Computing Systems

University Of Kentucky Research Foundation, Lexington KY

Investigators

Abstract

This Research Infrastructure Improvement Track-4 EPSCoR Research Fellows project will provide a fellowship to an Assistant professor and training for a graduate student at the University of Kentucky Research Foundation. This work will be conducted in collaboration with researchers at the Argonne National Laboratory (ANL). Message Passing Interface (MPI) is the de facto standard to perform communication and scale applications on high-performance computing systems. The performance of MPI is crucial to various downstream applications, including scientific simulations, big data analytics, and artificial intelligence. However, as the recent development of GPUs continues to outpace that of commodity networks, large-size data transfer is becoming the major performance bottleneck in state-of-the-art MPI libraries. This work aims to tackle this problem by developing a performant and scalable MPI library through integrated data compression, which is critical to fully exploit the power of current and next-generation computing systems. The success of this project will allow for accelerated executions of scientific code and data analytics, reducing the time to scientific insights for applications running on large-scale GPU-based computing systems. This will help advance scientific discoveries across a wide range of computer and computational disciplines. The deliverables of this project will be made publicly accessible to the community to enhance the research and engineering cyberinfrastructure in broader domains. In addition, this project will contribute to the education and workforce development for advanced cyberinfrastructure through the training of graduate students. The proposed project aims to deliver a high-performance and scalable compression-assisted MPI library to address the growing gap between the increasing computing power of GPU accelerators and relatively limited network bandwidth in high-end computing systems. The research and development work will be conducted at ANL through close collaborations with the leading experts in MPI and scientific data compression based on their established software products. Specifically, a composable GPU compression framework that features on-demand construction of compression pipeline will be developed first, in order to provide balanced trade-off between compression performance and message size reduction. This framework will then be leveraged to optimize the point-to-point communication in MPI. After that, tailored optimizations will be investigated for two important MPI collectives that are considered the major performance bottlenecks in scientific applications, and thorough error quantization will be performed through a combination of theoretical analysis and empirical evaluation. In addition, performance portability will be considered in the implementation to accommodate for the diverse architectures from different vendors. To this end, the developed routines will be integrated into the flagship MPI library MPICH (Message-Passing Interface Chameleon) and made publicly available to the research community. The evaluation of the deliverables will be performed on the leading computing facilities at ANL using two mission-critical scientific analyses. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.