Date of Award
Doctor of Philosophy
Jack Dongarra, George Bosilca, Michael Jantz, Yingkui Li
As the scale of High-performance Computing (HPC) systems continues to grow, researchers are devoted themselves to achieve the best performance of running long computing jobs on these systems. My research focus on reliability and efficiency study for HPC software.
First, as systems become larger, mean-time-to-failure (MTTF) of these HPC systems is negatively impacted and tends to decrease. Handling system failures becomes a prime challenge. My research aims to present a general design and implementation of an efficient runtime-level failure detection and propagation strategy targeting large-scale, dynamic systems that is able to detect both node and process failures. Using multiple overlapping topologies to optimize the detection and propagation, minimizing the incurred overhead sand guaranteeing the scalability of the entire framework. Results from different machines and benchmarks compared to related works shows that my design and implementation outperforms non-HPC solutions significantly, and is competitive with specialized HPC solutions that can manage only MPI applications.
Second, I endeavor to implore instruction level parallelization to achieve optimal performance. Novel processors support long vector extensions, which enables researchers to exploit the potential peak performance of target architectures. Intel introduced Advanced Vector Extension (AVX512 and AVX2) instructions for x86 Instruction Set Architecture (ISA). Arm introduced Scalable Vector Extension (SVE) with a new set of A64 instructions. Both enable greater parallelisms. My research utilizes long vector reduction instructions to improve the performance of MPI reduction operations. Also, I use gather and scatter feature to speed up the packing and unpacking operation in MPI. The evaluation of the resulting software stack under different scenarios demonstrates that the approach is not only efficient but also generalizable to many vector architecture and efficient.
Zhong, Dong, "Toward Reliable and Efficient Message Passing Software for HPC Systems: Fault Tolerance and Vector Extension. " PhD diss., University of Tennessee, 2021.