Research On Quality Of Service Guarantee Of High Performance Computing Storage System Based On SSD

With the expansion of high-performance computing(HPC)clusters and the growth of computing performance,contention for storage system IO resources are becoming more and more frequent.It is becoming more and more important to provide service quality assurance for applications in the HPC cluster environment.Different from the traditional network field,applications in the HPC environment pay more attention to guaranteeing the lower limit of IO resources.Now in the research of service quality reservation guarantee in the storage field,it is mainly implemented through an exclusive resource scheduling scheme or an algorithm based on priority queues,but these schemes are designed for IO resource management of virtual machines in a cloud environment.It is impossible to provide an accuracy guarantee for the lower limit of IO resources in the HPC cluster environment.In order to solve these problems,SDR-QoS,a dynamic guarantee method for QoS reservation in SSD-based distributed storage systems,is proposed to guarantee QoS reservation by dynamically modifying application QoS limits.The distributed storage system based on SSD not only can provide stable aggregate bandwidth,but its bandwidth distribution also related to the number of clients and IO block size.First,SDR-QoS allocates application's QoS requirements to specific storage servers based on the bandwidth prediction model.Then,SDR-QoS tags the IO request on the storage server and processes the request according to the tag value.In addition,under the guidance of the QoS dynamic adjustment strategy,SDR-QoS divides the stable aggregate bandwidth provided by the SSD according to the demand allocated by the application,and dynamically sets the upper limit of the bandwidth of the low-priority application to reduce the preemption of the priority queue,thereby ensuring the application QoS reservation.Experimental results show that in multi-application concurrent scenarios,SDR-QoS can not only accurately guarantee QoS reservations that exceed 90% of the total bandwidth of the service node,but also accurately and dynamically adjust to changing QoS requirements.At the same time,the performance loss caused can also be ignored.