On the assessment of probabilistic WCET estimates reliability for arbitrary programs

Suzana Milutinovic,Francisco J Cazorla,Jaume Abella

doi:10.1186/s13639-017-0076-8

Abstract

Measurement-Based Probabilistic Timing Analysis (MBPTA) has been shown to be an industrially viable method to estimate the Worst-Case Execution Time (WCET) of real-time programs running on processors including several high-performance features. MBPTA requires hardware/software support so that program’s execution time, and so its WCET, has a probabilistic behaviour and can be modelled with probabilistic and statistic methods. MBPTA also requires that those events with high impact on execution time are properly captured in the (R) runs made at analysis time. Thus, a representativeness argument is needed to provide evidence that those events have been captured.This paper addresses the MBPTA representativeness problems caused by set-associative caches and presents a novel representativeness validation method (ReVS) for cache placement. Building on cache simulation, ReVS explores the probability and impact (miss count) of those cache placements that can occur during operation. ReVS determines the number of runs R′, which can be higher than R, such that those cache placements with the highest impact are effectively observed in the analysis runs, and hence, MBPTA can be reliably applied to estimate the WCET.

Highlights

The validation and verification (V&V) process for critical real-time systems requires collecting sufficient evidence that critical functions will execute correctly and timely
We present Representativeness Validation by Simulation (ReVS), a method valid for arbitrary cache access patterns to assess whether probabilistic WCET (pWCET) estimates obtained with Measurement-Based Probabilistic Timing Analysis (MBPTA)—for a given number of runs—are reliable
8 Conclusions MBPTA uses EVT to estimate the pWCET of programs

Summary

Introduction

The validation and verification (V&V) process for critical real-time systems requires collecting sufficient evidence that critical functions will execute correctly and timely. The execution time of those runs in which the number of addresses (randomly) mapped to a cache set exceeds its associativity (W ) can be significantly higher than when this is not the case [10] This fact becomes an issue for MBPTA when those cache placements of interest occur with a sufficiently high probability to be deemed as relevant by the corresponding safety standard (e.g. above 10−9), but sufficiently low not to be observed in the measurements at analysis time (e.g. below 10−3) [10,11,12].

Problem statement

ReVS main steps ReVS includes the following steps

Combined aCi impact and probability

ReVS results

Assessing ReVS reliability

Conclusions