Demystifying Power and Performance Variations in Gpu Systems Through Microarchitectural Analysis

Abstract

Graphics Processing Units (GPUs) serve efficient parallel execution for general-purpose computations at high-performance computing and embedded systems. While performance concerns guide the main optimization efforts, power issues become significant for energy-efficient and sustainable GPU executions. Profilers and simulators report statistics about the target execution; however, they either present only performance metrics in a coarse kernel function level or lack visualization support that can enable microarchitectural performance analysis or performance-power consumption comparison. Evaluating runtime performance and power consumption dynamically across GPU components enables a comprehensive tradeoff analysis for GPU architects and software developers. In this work, we present a novel memory performance and power monitoring tool for GPU programs, GPPRMon, which performs a systematic metric collection and provides useful visualization views to guide power and performance analysis for target executions. Our simulation-based framework dynamically gathers SM and memory-related microarchitectural metrics by monitoring individual instructions and reports dynamic performance and power values. Our interface presents spatial and temporal views of the execution. While the first demonstrates the performance and power metrics across GPU memory components, the latter shows the corresponding information at the instruction granularity in a timeline. We demonstrate performance and power analysis for memory-bound graph applications and resource-critical embedded programs from GPU benchmark suites. Our case studies reveal potential usages of our tool in memory-bound kernel identification, performance bottleneck analysis of a memory-intensive workload, performance-power evaluation of an embedded application, and the impact of input size on the memory structures of an embedded system.