HomePublicationsPaper

Measurement-Based Probabilistic Timing Analysis for Graphics Processor Units
Ref: CISTER-TR-160302       Publication Date: 4 to 7, Apr, 2016

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.

Kostiantyn Berezovskyi

,

Fabrice Guet

,

Luca Santinelli

,

Konstantinos Bletsas

,

Eduardo Tovar

Document:

View
PDF Download
PDF