Oversubscription of computational resources on multicore desktop systems

Authors:
Constantin Christmann;Erik Hebisch;Anette Weisbecker
Affiliations:
Institut für Arbeitswissenschaft und Technologiemanagement, Universität Stuttgart, Stuttgart, Germany;Institut für Arbeitswissenschaft und Technologiemanagement, Universität Stuttgart, Stuttgart, Germany;Institut für Arbeitswirtschaft und Organisation IAO, Fraunhofer, Stuttgart, Germany
Venue:
MSEPT'12 Proceedings of the 2012 international conference on Multicore Software Engineering, Performance, and Tools
Year:
2012

Citing 8
Cited 0

Dynamically Adapting the Degree of Parallelism with Reflexive Programs

IRREGULAR '96 Proceedings of the Third International Workshop on Parallel Algorithms for Irregularly Structured Problems
Parallel Parameter Tuning for Applications with Performance Variability

SC '05 Proceedings of the 2005 ACM/IEEE conference on Supercomputing
Automatic tuning of whole applications using direct search and a performance-based transformation system

The Journal of Supercomputing
Patterns for parallel programming

Patterns for parallel programming
Intel threading building blocks

Intel threading building blocks
Auto-tuning support for manycore applications: perspectives for operating systems and compilers

ACM SIGOPS Operating Systems Review
Windows Internals: Book and Online Course Bundle

Windows Internals: Book and Online Course Bundle
A case for machine learning to optimize multicore performance

HotPar'09 Proceedings of the First USENIX conference on Hot topics in parallelism

Quantified Score

Hi-index	0.00

Visualization

Abstract

The increasing pervasiveness of multicore processors in today's computing systems will increase the demand for techniques to adapt application level parallelism. In this paper we demonstrate the impact of oversubscription in the context of a desktop scenario and show that using too many threads in one application can lead to significant performance loss for other applications within the system environment (in our example more than 30%). We examine how common parallelization tools are trying to exploit parallelism while not causing oversubscription and find that common tools do not offer a mechanism which sufficiently takes into account the computational load of the system. Oversubscription is not easy to detect from an application perspective due to the fair scheduling of common operating systems. Nonetheless, we do present an approach for detecting oversubscription by exploiting a scheduling artifact in the average execution time of a parallel code section.