Web proceedings papers

Authors

Genti Daci and Megi Tartari

Abstract

Contention for shared resources on multicore processors is an emerging issue of great concern, as it affects directly performance of multicore CPU systems. In this regard, Contention-Aware scheduling algorithms provide a convenient and promising solution, aiming to reduce contention, by applying different thread migration policies to the CPU cores. The main problem faced by latest research when applying these schedulers in different multicore systems, was a significant variation of performance achieved on different system architectures. We aim to review and discuss the main reasons of such variance arguing that most of the scheduling solutions were designed based on the assumption that the underlying system was UMA (Uniform Memory Access latency, single memory controller), but modern multicore systems are NUMA (Non Uniform Memory Access latencies, multiple memory controllers). Thispaper focuses on reviewing the challenges on solving the contention problem for both types of system architectures. In this paper, we also provide a comparative evaluation of the solutions applicable to UMA systems which are the most extensively studied today, discussing their features, strengths and weaknesses. For addressing performance variations, we will review Vector Balancing, OBS-X and DIO scheduling for UMA systems. While for NUMAsystems, we will compare and discuss DINO and AMPS Schedulers which supports NUMA architectures aiming to resolve performance issues and also introduce the problems they have. This paper aims to propose further improvements to these algorithms aiming to solve more efficiently the contention problem, considering that performance-asymmetric architectures may provide a cost-effective solution.

Keywords

Uniform Memory Access(UMA), Multicore CPU systems, Contention-Aware Scheduling, Non Uniform Memory Access(NUMA), Vector Balancing Scheduling, OBS-X Scheduler, DIO Scheduler, DINO Scheduler, AMPS Scheduler.