Scheduling techniques to avoid contention in multi-core systems

摘要

One of the main problems in multi-core systems is the contention of shared resources such as cache, memory controller, pre-fetcher etc. among the cores. Due to the contention among shared resources, the processing unit's performance is degraded. Scheduling of applications in such a way that it reduces the contention among shared resources is one of the promising solutions. Scheduling is considered as an efficient and best technique as it doesn't require any extra hardware or any changes to be made to the OS or its underlying kernel. Scheduling can be implemented at user level by using system calls. In the prior works it was considered that the cache contention was the main cause of performance degradation and many hardware and software techniques were found to avoid or minimize it. But further experiments proved that the contention caused by pre-fetcher and memory controller is also having significant effect on performance degradation. Many scheduling policies and classification schemes have been designed to find out an efficient scheduling algorithm. Miss rate is considered to be simple yet efficient classification scheme to classify the threads as it not only considers contention due to cache but also the memory controller and pre-fetcher. Distributed Intensity is the first scheduling algorithm discussed which uses miss rate to classify threads and assign them to all cores in an efficient way so that miss rate is shared almost equally among the cores. Then Distributed Intensity is combined with Swap algorithm to further improve the performance by using dynamic optimization. Then by further studies it is found out that miss rate cant be efficient classification technique for memory intensive workloads. So the concepts of Contentiousness and Sensitivity are introduced to improve the efficiency of scheduling algorithm and to minimize the performance degradation due to contention.