Characterization of Shared Resource Contention in Multi-core Systems

(1)

Characterisation of Shared Resource

Contention in Multi-core Systems

Jakob Danielsson

Ja ko b D a n ie lss o n C HA R A C TE R IS A TI O N O F S HA R ED R ES O U RC E C O N TE N TI O N I N M U LTI -C O R E S YS TE M S 20 19

Mälardalen University Licentiate Thesis 287

ISBN 978-91-7485-449-7 ISSN 1651-9256

Address: P.O. Box 883, SE-721 23 Västerås. Sweden Address: P.O. Box 325, SE-631 05 Eskilstuna. Sweden E-mail: info@mdh.se Web: www.mdh.se

Jakob Danielsson is a PhD student at Mälardalen University since August, 2016. Jakob received his B.Sc. degree in Computer Science from Mälardalen University, Sweden in 2014. In 2016, Jakob obtained a Masters degree in embedded systems. The same year in August, Jakob Started his PhD studies at Mälardalen University. His research interests include workload characterization, shared resource contention and resource isolation techniques in multi-core systems.

Multi-core computers are infamous for being hard to use in time-critical systems due to execution-time variations as an effect of shared resource contention. In this thesis we study the problem of shared resource contention which occurs when multiple applications executing on different cores and do not have exclusive ownership of a shared resource. We investigate performance variations of parallel tasks in multi-core systems and pinpoint the source of the performance variation and resource contention using performance counters. Furthermore, we investigate methods to mitigate performance variations using resource isolation techniques. We present a methodology for verifying isolation and tested the achieved isolation using the Jailhouse hypervisor. We further investigate shared cache memory isolation opportunities using a page coloring tool called PALLOC. Page-coloring is used for partitioning the cache, assigning specific cache lines to specific processes. Page coloring can however cause system performance degradation since it decreases the to-tal amount of cache memory available for each process. Finally, we propose a dynamic partitioning assignment policy which assigns cache partitions to a process according to an adaptive model based on the process performance. The general conclusion from our investigations is that a large body of applications can suffer from shared resource contention and that techniques for mitigating resource contention. Our methods measure characterise applications, identifies resource contention and finally suggests on isolation techniques.

(2)

Mälardalen University Press Licentiate Theses No. 287