Multicore Performance Prediction – Comparing Three Recent Approaches in a Case Study

verfasst von

Matthias Lüders, Oliver Jakob Arndt, Holger Blume

Abstract

Even though parallel programs, written in high-level languages, are portable across different architectures, their parallelism does not necessarily scale after migration. Predicting a multicore-application’s performance on the target platform in an early development phase can prevent developers from unpromising optimizations and thus significantly reduce development time. However, the vast diversity and heterogeneity of system-design decisions of processor types from HPC and desktop PCs to embedded MPSoCs complicate the modeling due to varying capabilities. Concurrency effects (caching, locks, or bandwidth bottlenecks) influence parallel runtime behavior as well. Complex performance prediction approaches emerged, which can be grouped into: virtual prototyping, analytical models, and statistical methods. In this work, we predict the performance of two algorithms from the field of advanced driver-assistance systems in a case study. With the following three methods, we provide a comparative overview of state-of-the-art predictions: GEM5 (virtual prototype), IBM Exabounds (analytical model), and an in-house developed statistical method. We first describe the theoretical background, describe the experimental- and model-setup, and give a detailed evaluation of the prediction. In addition, we discuss the applicability of all three methods for predicting parallel and heterogeneous systems.

Organisationseinheit(en)

Institut für Mikroelektronische Systeme

Typ

Aufsatz in Konferenzband

Seiten

282-294

Anzahl der Seiten

13

Publikationsdatum

2020

Publikationsstatus

Veröffentlicht

Peer-reviewed

Ja

ASJC Scopus Sachgebiete

Theoretische Informatik, Allgemeine Computerwissenschaft

Elektronische Version(en)

https://doi.org/10.1007/978-3-030-48340-1_22 (Zugang: Geschlossen)