Bounding the WCET of a GPU Thread Block with a Multi-Phase Representation of Warps Execution (Artifact)

Abstract

This paper proposes to model the Worst-Case Execution Time (WCET) of a GPU thread block as the Worst-Case Response Time (WCRT) of the warps composing the block. Inspired by the WCRT analyzes for classical CPU tasks, the response time of a warp is modeled as its execution time in isolation added to an interference term that accounts for the execution of higher priority warps. We provide an algorithm to build a representation of the execution of each warp of a thread block that distinguishes phases of execution on the functional units and phases of idleness due to operations latency. A simple formula relying on this model is then proposed to safely upper bound the WCRT of warps scheduled under greedy policies such as Greedy-Then-Oldest (GTO) or Loose Round-Robin (LRR). We experimented our approach using simulations of kernels from a GPU benchmark suite on the Accel-Sim simulator. We also evaluated the model on a GPU program that is likely to be found in safety critical systems : SGEMM (Single-precision GEneral Matrix Multiplication). This work constitutes a promising first building block of an analysis pipeline for enabling static WCET computation on GPUs.