In-Depth Analyses of Unified Virtual Memory System for GPU Accelerated Computing

The abstraction of a shared memory space over separate CPU and GPU memory domains has eased the burden of portability for many HPC codebases. However, users pay for the ease of use provided by systems-managed memory space with a moderate-to-high performance overhead. NVIDIA Unified Virtual Memory (UVM) is presently the primary real-world implementation of such abstraction and offers a functionally equivalent testbed for a novel indepth performance study for both UVM and future Linux Heterogeneous Memory Management (HMM) compatible systems. The continued advocation for UVM and HMM motivates the improvement of the underlying system. We focus on a UVM-based system and investigate the root causes of the UVM overhead, which is a non-trivial task due to the complex interactions of multiple hardware and software constituents and the requirement of targeted analysis methodology. In this paper, we take a deep dive into the UVM system architecture and the internal behaviors of page fault generation and servicing. We reveal specific GPU hardware limitations using targeted benchmarks to uncover driver functionality as a real-time system when processing the resultant workload. We further provide a quantitative evaluation of fault handling for various applications under different scenarios, including prefetching and oversubscription. We find that the driver workload is dependent on the interactions among application access patterns, GPU hardware constraints, and Host OS components. We determine that the cost of host OS components is significant and present across implementations, warranting close attention. This study serves as a proxy for future shared memory systems such as those that interface with HMM.