Effective Quality-of-Service Policy for Capacity High-Performance Computing Systems

摘要

Current high-performance capacity systems run hundreds of applications concurrently sharing the same network resources. In this scenario, system throughput and hence system utilization is significantly affected by inter-application contention. Quality-of-service (QoS) mechanisms are traditionally used in the Internet to provide different levels of performance guarantees. However, we have not found in the literature the application of QoS techniques to reduce the impact of inter-application contention on contention-sensitive applications running in capacity high-performance computing (HPC) clusters. Only a small set of applications can benefit from the policies currently used. Applications are basically classified into two classes, namely latency-sensitive and bandwidth-sensitive applications. Instead, we propose a more efficient quality-of-service policy for HPC systems with InfiniBand network. This technique provides a wider classification of applications, and hence can leverage QoS mechanisms for a large number of applications resulting in an increase of system performance. The proposed QoS policy provides a method to effectively map applications to virtual lanes and applies effective bandwidth distribution for each of the lanes. Results obtained for slimmed fat-tree networks show that by segregating applications on virtual lanes a substantial reduction of inter-application contention by 59% can be achieved for typical scientific codes. Also, applications can be effectively grouped into one virtual lane based on a network utilization metric in case there are more applications than virtual lanes in the system. Moreover, further improvements are obtained by allocating a different bandwidth to each virtual lane. This technique will favor some applications without significantly degrading the others. We show that for HPC, contrary to Internet, QoS policies should favor applications that are less communication demanding.