Holistic Performance Analysis and Optimization of Unified Virtual Memory

Author: Tyler Allen (Clemson University)

Advisor: Rong Ge (Clemson University)

Abstract: High-performance computing systems have seen tremendous growth in theoretical performance with the inclusion of Graphics Processing Units (GPUs) and other accelerators. The difficulty of programming these systems has grown alongside the performance as programmers are required to manage separate programming models for accelerators alongside traditional CPU resources. This complexity was mitigated with the introduction of heterogeneous shared memory systems, such as NVIDIA Unified Virtual Memory (UVM) and Heterogeneous Memory Management (HMM), to replace the requirement of programmers to directly manage memory. These technologies reduce programming complexity by managing the physical location of memory on behalf of the programmer. However, there is a significant performance gap between heterogeneous shared memory and directly-managed memory in modern systems, significantly reducing effective application performance and the efficiency of the underlying HPC systems. While prior work has taken steps to understand the high-level performance of these systems, the underlying performance issues are not well understood and therefore challenging to resolve. In this work, we introduce our methodology for deep investigation of UVM performance at the systems-software and workload-generation level. This work will identify the fundamental costs within UVM, including the fundamental costs of UVM systems software as well as the sources of overhead that can be eliminated. Based on this initial performance study, we also investigate optimizations to UVM to increase overall system performance and improve HPC application performance automatically. Further, we seek to generalize these improvements and insights such that they can be applied to other systems software and devices, such as HMM.

Thesis Canvas: pdf

Back to Doctoral Showcase Archive Listing