Performance Oriented Resource Scheduling in Cloud Storage Systems

Zhihao Yao, Purdue University


Achieving high performance in cloud storage systems and virtualized data centers requires both a high-end storage infrastructure and an intelligent resource scheduling algorithm to make the best use of hardware resources. The current scheduling algorithm design only takes capacity and backend properties into account to make a scheduling decisions. Hence cloud service providers are not able to effectively manage the service performance and offer a competitive and assurable performance SLA. In fact, the scheduling algorithm in a widely deployed cloud platform, OpenStack, can cause resource contention and performance degradation in the storage cluster. In this research, we addressed the storage resource scheduling problem for the cloud block storage systems. We first proposed two performance SLA oriented scheduling algorithms that enabling performance management at resource scheduling layer and minimize the SLA violation rate. The first SLA oriented scheduling algorithm takes IOPS performance into account and prioritizes candidate hosts based on their available IOPS resource. Our evaluation results show that our proposed algorithm is able to reduce 20% of SLA violations and higher volume I/O throughput with fewer storage hosts. The second multi-dimensional scheduling algorithm considers multiple resource requirements during scheduling and guarantees performance SLA with the dynamic workload. To solve the resource scheduling and workload consolidation problem in an SSD based block storage system, we first rigorously investigated the latency performance characteristics of four enterprise SSDs by applying a linear regression analysis. We found that there is a correlation between workload parameters and aggregated host latency performance measured in average and at the 95th percentile. Based on the performance model, we designed a static scheduling algorithm to make latency optimized placement decisions for new workload requests. Next, we developed a dynamic load balancing algorithm to further optimize the global latency. The evaluation results show the effectiveness of our proposed scheduling strategy.




Papapanagiotou, Purdue University.

Subject Area

Computer Engineering|Information Technology

Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server