| The explosive growth of the Internet and applications has been driving for high performance, scalable and reliable storages for data access. As a result of this, on one hand, network-enabled storage such as NAS, SAN, iSCSI and OSD are emerging and becoming popular. On the other hand, the storage devices are getting more intelligent. These new storage paradigms have created new design issues. In this thesis, we investigate two relaxed issues: performance and QoS provisioning of such infrastructure.;The performance issue is becoming important since the storage access is different from traditional data access for which the underlying networking protocol was designed. In this thesis, we first study the performance characteristics of iSCSI-based storage systems under different configurations and investigate the system bottleneck and the factors affecting the performance. Moreover, to facilitate the study of this emerging technology, we also investigate the modeling and simulation of iSCSI storage systems. We exploit the existing ns2 tool and incorporate these modeling to support iSCSI-based storage simulation. We validate this tool with the real measurement results. We also investigate the impact of several transport factors and iSCSI parameters to illustrate the usage of this tool.;We then examine the Quality of Service (QoS) issue, i.e. how to provide QoS support to the clients and client's applications in the networked storage system. QoS can provide certain service guarantees for the requests from clients. It also supports service differentiation among requests from different clients and applications. We first analyze the QoS requirements from a client applications' perspective, and then propose extensions to OSD protocol to incorporate the specification of QoS requirements.;We further work on the enforcement of QoS in the remote storage device. We propose a novel scheduling scheme to address the QoS requirements of client applications. In such scheme, we consider both the session's network bandwidth and the storage conditions. We also break down requests of large size into smaller size of appropriate size and schedule the requests based on network condition, current workload and request's deadline. In addition, we have also implemented such prototyping system and experimentally demonstrated its efficacy. |
