Research On Storage Quality Of Service For Cloud Data Centers

Nowadays the rapid growth of internet applications and data volumes is pressuring storage systems toward a service-oriented paradigm, in which numerous customers share storage resources to bring down the cost of managing and deploying storage systems. With such a consolidation of storage resources, customers purchase storage service based on the amount of storage space and performance of accessing their data. However, applications sharing this storage infrastructure usually compete for storage resource and interfere with each other, which may lead to great challenges of Quality of Service (QoS) guarantees and storage performance optimization, especially in complicated application environments. In such environment, applications exhibit diversification and high bursty, and storage systems get more complexity and heterogeneity. Therefore, how to allocate storage resource for providing Quality of Service guarantees while enhancing merit of data centers has attracted significant attention.In service-oriented environment, concurrent applications running on shared storage system usually has different performance requirement and exhibit high bursty. For adapt-ing such environment, a novel scheduling framework (Courier) is designed to provide per-application performance guarantees and enhance storage performance. Courier logically interleaves a feedback-based latency controller and reward credit-based scheduler in fine-grained manner. The latency controller restricts I/O response time of each request through a feedback-based mechanism and handling urgent request preferentially. When all requests are expected to meet their latency requirement, Courier invoke the reward budget-based al-gorithm, which allows applications overdraw their future credits for burst urgent requests and rewards well-behavior applications with more spare storage resource to achieve high storage utilization. Benefitting from cooperation of feedback-based latency controller and reward credit-based mechanism, Courier can meet per-application performance targets in terms of both throughput and latency, while yield about20%performance improvement.To provide performance isolation while maintaining high I/O efficiency for applica- tions on Solid State Drive(SSD), a2-level I/O scheduler (FBCQ) is proposed to exploit internal architecture and performance characteristics of SSD. The higher level scheduler of FBCQ manages performance of concurrent applications in terms of service times, and dis-tinguish read and write requests based on the performance characteristics of SSD. At the lower level, channel-based queues are designed to exploit the inherent channel-level paral-lelism of SSD. Each request is placed into corresponding channel-based queue based on its physical address, and then, channel-based queues dispatch their outstanding requests simul-taneously. Within each channel-based queue, scheduler gives higher priority to read requests than writes to mitigate read-blocked-by-write interference and improves write performance through sorting and bundling write requests. Experiment results show the2-level scheduler outperforms other schedulers at both aspects of performance isolation and I/O efficiency.State-of-the-art SSD and conventional hard disk has both strengths and limitations in aspects of latency, cost and capability. Therefore, the hybrid storage systems that combine hard disk and SSD can alleviate limitations and exploit their advantages. In such environ-ment, concurrent applications compete for flashcache and interfere with each other. To over-come this problem, the QoS-aware flash cache allocation (QoS-FCA) scheme is proposed to meet latency targets for applications while improve performance of hybrid storage systems. Based on the analysis of access protocol in flashcache, pre-application performance model is built to describe the relationship between pre-application latency and its flashcache size. Thus, the QoS-FCA can allocate flash cache to applications based on their performance model and workload characteristic dynamically. The results show that QoS-FCA is able to provide latency requirement for applications while improving hit rate as much as about10%.The popularity of internet applications and scientific computation strongly urges underlying storage systems to be capable of managing billions of small files and providing high metadata performance. The analysis of file access protocol has shown that com-munications for metadata access consumes the overwhelming majority of one individual file access. Based on this observation, a delegable metadata service (DMS) is built in client-side memory to absorb metadata requests. To cooperate with DMS, mlock is designed to maintain consistency for file system. In addition, mdflush is implemented in DMS to group multiple metadata updates into a single message and perform metadata updates at background for hiding metadata access latency. When compared to original Cappella file system, the DMS scheme can yield more than35%performance improvement for small-file accesses.