Performance Management In Storage Systems

With the booming development of computer theories and technologies and the endless springing up of new high-end applications, practical requirements drive the size, scale, and complexity of storage systems exploding, and thus the cost of storage systems management has held a big piece of the pie of total system cost. Therefore, storage systems management becomes a crucial, urgent, imperative task in the storage systems domain.As an important area in the field of storage systems management, performance management has attracted much attention from the computer systems community. The essential issues in the performance management area primarily include quality-of-service (QoS) guarantees (mainly, performance guarantees), system configuration management, performance problem diagnosis, performance modeling, capacity planning, performance tuning, and so on. Centered on the four core problems in the performance management area, this dissertation has promoted the techniques addressing the problems. The contributions as follows:To decrease the complexity of storage systems management, facilitate systems management, and reduce system operations cost, shared storage servers are extensively deployed in product environments. However, storage resource sharing results in interference among workloads, and the performance of individual competing application becomes unpredictable. We present an I/O scheduling algorithm for shared storage servers in the first part of the thesis. The design philosophy of the algorithm is Weighted Complete Fair Queuing (WCFQ). Based on the idea, the algorithm conducts the allocation of underlying storage resources among multiple concurrent workloads and consequently fulfills performance guarantees of individual workload.The storage consolidation paradigm is widely adopted in today's large-scale data centers, whether in virtualized storage servers, or in network-attached storage systems, or in cluster storage systems. To accomplish the efficient performance management of applications while improving the utilization of storage systems, we present a utilization-driven performance management framework that, due to the dynamics of both of storage systems and workloads, is designed with the adaptive control methodology. Complying with the prescribed utilization objective, the performance management system periodically captures the system model online and dynamically adjusts the parameters of the controller. The actuator dictated by the controller takes proper control actions to achieve the performance management goals. In addition, for the purpose of exploring, we detail the design of performance management systems respecting the relative quality-of-service model. Both of the performance management systems are depicted in the second part of the thesis.Performance models of storage systems are fairly valuable for storage systems design, systems management, and performance evaluation. Differing from the traditional first-principles modeling techniques, we propose a black box-style approach for performance model extraction of storage systems in the third part of the thesis. The approach first monitors the characteristics of I/O workloads and the system performance and next, leveraging the multivariable regression theory, figures out the mapping relationship between both of the workload profile and the performance measures (i.e., performance model extraction). The experimental results demonstrate that the performance model derived from our approach can achieve rather high accuracy. Also, we validate the practicability and effectiveness of the applications of the performance model to the problems of performance prediction and what-if analysis, both of which are the core tasks in the domain of storage systems management.Last but not least, the theme of the fourth part of the thesis is about performance tuning. Both of the large scale and high complexity of high performance storage systems make the system configuration increasingly difficult, and the default configuration, or empirical configuration of a storage system usually induces sub-optimal, even pathological system performance. Performance tuning can not only significantly increase performance of applications but also improve utilization of system resources saliently and thus reduce total system cost. To this end, we present a statistical analysis-based approach to performance tuning in storage systems. It consists of two phases:critical system parameters identification and critical parameters performance optimization. In the first phase we employ the analysis-of-variance (ANOVA) method to identify the critical system parameters that have a significant effect on the application performance. On the basis of the previous stage, we use the response surface method (RSM) to conduct the performance optimization analysis of the identified knobs. The task of the second phase is to find the niche combination of the critical system parameters that can achieve the optimal performance and eventually fulfill the objective of performance tuning.