On mitigating memory bandwidth contention through bandwidth-aware scheduling

Abstract

Shared-memory multiprocessors have dominated all platforms from high-end to desktop computers. On such platforms, it is well known that the interconnect between the processors and the main memory has become a major bottleneck. The bandwidth-aware job scheduling is an effective and relatively easy-to-implement way to relieve the bandwidth contention. Previous policies understood that bandwidth saturation hurt the throughput of parallel jobs so they scheduled the jobs to let the total bandwidth requirement equal to the system peak bandwidth. However, we found that intra-quantum fine-grained bandwidth contention still happened due to a program's irregular fluctuation in memory access intensity, which is mostly ignored in previous policies.In this paper, we quantify the impact of bandwidth contention on overall performance. We found that concurrent jobs could achieve a higher memory bandwidth utilization at the expense of super-linear performance degradation. Based on such an observation, we proposed a new workload scheduling policy. Its basic idea is that interference due to bandwidth contention could be minimized when bandwidth utilization is maintained at the level of average bandwidth requirement of the workload. Our evaluation is based on both SPEC 2006 and NPB workloads. The evaluation results on randomly generated workloads show that our policy could improve the system throughput by 4.1% on average over the native OS scheduler, and up to 11.7% improvement has been observed.

Contention-Aware Scheduling for SMT Multicore Processors

“…This fact complicates performing an adequate evaluation of scheduling algorithms. For instance, Xu et al [11] observed that a scheduling policy that prioritizes the longest jobs could provide the best turnaround time in most workloads when the benchmarks experience widely different execution times. Another important drawback is that benchmarks with different execution time will have different weights in the mix execution, which might not be correctly reflected in some performance and fairness evaluation metrics.…”

“…However, targeting fairness and performance at the same time is not an easy task. For example, a prevalent approach to improve performance consists in balancing the memory requests of a multiprogram workload along its execution time [11] [12]. In contrast, to improve fairness, the processes with less accumulated progress should be prioritized over processes with superior progress.…”

“…More recently, Xu et al [11] prove that irregular memory access patterns can produce fine-grained contention when the required bandwidth is close to the peak bandwidth.…”

Mitigating Memory Resource Contention in Warehouse Scale Computers

“…System software also does not have control over hardware resources such as caches and memory bandwidth, which renders responding to contention and managing applications' QoS quite challenging. As a result, despite the amount of research attention given to contention problems on multicore platforms [28,45,10,19,44,51,52,27,32,15,29,24,63,23,24,7,4,60], mitigating the impact of contention on an application's performance and QoS, enforcing the relative QoS priorities of co-running applications, while maximizing machine utilization, remains key challenges in modern warehouse scale computers.…”

“…[Contention Aware Scheduling] A research direction that recently attracted a wealth of attention is contention aware scheduling [15,29,24,63,23,24,7,4,60]. Contention aware scheduling techniques use predictors or models to decide what applications should be co-running together to minimize the performance degradation or to improve performance isolation.…”