Measurement-Based Probabilistic Timing Analysis for Graphics Processor Units

Purely analytical worst-case execution time (WCET) estimation approaches for Graphics Processor Units (GPUs) cannot go farbecause of insufficient public information for the hardware. Thereforemeasurement-based probabilistic timing analysis (MBPTA) seems theway forward. We recently demonstrated MBPTA for GPUs, based onExtreme Value Theory (EVT) of the “Block Maxima” paradigm. In thisnewer work, we formulate and experimentally evaluate a more robustMBPTA approach based on the EVT “Peak over Threshold” paradigmwith a complete set of tests for verifying EVT applicability. It optimallyselects parameters to best-fit the input measurements for more accurateprobabilistic WCET estimates. Different system configuration parameters (cache arrangements, thread block size) and their effect on thepWCET are considered, enhancing models of worst-case GPU behavior. Measurement-Based Probabilistic Timing Analysis for Graphics Processor Units Kostiantyn Berezovskyi, Fabrice Guet, Luca Santinelli, Konstantinos Bletsas, and Eduardo Tovar ∗ONERA Toulouse, France, CISTER/INESC-TEC, ISEP, Portugal. Abstract. Purely analytical worst-case execution time (WCET) estimation approaches for Graphics Processor Units (GPUs) cannot go far because of insufficient public information for the hardware. Therefore measurement-based probabilistic timing analysis (MBPTA) seems the way forward. We recently demonstrated MBPTA for GPUs, based on Extreme Value Theory (EVT) of the “Block Maxima” paradigm. In this newer work, we formulate and experimentally evaluate a more robust MBPTA approach based on the EVT “Peak over Threshold” paradigm with a complete set of tests for verifying EVT applicability. It optimally selects parameters to best-fit the input measurements for more accurate probabilistic WCET estimates. Different system configuration parameters (cache arrangements, thread block size) and their effect on the pWCET are considered, enhancing models of worst-case GPU behavior. Purely analytical worst-case execution time (WCET) estimation approaches for Graphics Processor Units (GPUs) cannot go far because of insufficient public information for the hardware. Therefore measurement-based probabilistic timing analysis (MBPTA) seems the way forward. We recently demonstrated MBPTA for GPUs, based on Extreme Value Theory (EVT) of the “Block Maxima” paradigm. In this newer work, we formulate and experimentally evaluate a more robust MBPTA approach based on the EVT “Peak over Threshold” paradigm with a complete set of tests for verifying EVT applicability. It optimally selects parameters to best-fit the input measurements for more accurate probabilistic WCET estimates. Different system configuration parameters (cache arrangements, thread block size) and their effect on the pWCET are considered, enhancing models of worst-case GPU behavior.