Research On End To End QoS Support Strategy In ISCSI OSD Storage System

With the development of network storage intelligence, OBNS(Object Based Network Storage) joined technique advantage of both the block-level interface to the SAN and the file-level interface to the NAS is generated. It has high intelligence, high performance, high reliability, cross-platform, as well as the safety of data sharing characteristic and so on. And domestic and international research of storage QoS (Quality of Service) in this area began shortly. It is also unable to meet the increasing seriously storage requirements in the breadth and depth, and furthermore storage QoS has not formed an effective architecture. Therefore storage QoS is a very important research direction in research of network storage technology.iSCSI OSD (Object Storage Device) storage systems is a network storage technology based on object, which is needed to extend the iSCSI protocol for the OSD protocol. Based on the research of iSCSI OSD architecture, it adopts CQNM (Closed Queuing Network Model) to analysis the performance interval of iSCSI OSD storage system, and sets up model formula for max and min of throughput and response time. And it can be used as the judgement for the rationality of clients requirement for the storage QoS.In iSCSI OSD it provides end-to-end QoS support between client and storage system from four aspects, including object attributes with the QoS support, dynamic network bandwidth allocation with QoS support, dynamic cache allocation with QoS support and storage device with the QoS support. Object attributes with the QoS support is implemented by extending Object attributes for Bandwidth,ResponseTime and IOPS in OSD protocol. Dynamic network bandwidth allocation with QoS support is implemented by calculating network bandwidth needed according to network protocol overhead and storage QoS requirements of clients, and adopting a dynamic network bandwidth allocation strategy. According to the relationship between the hitrate in cache and the responsetime, dynamic cache allocation with QoS support is implemented by a dynamically allocating buffer strategy to meet the responsetime requirements of different clients. It adopts hiberarchy I/O scheduling of QoS support in storage device, which are object I/O scheduling of QoS support and disk I/O scheduling of QoS support. Object I/O scheduling of QoS support includes object I/O pretreatment, establishing QoS queue and calculating the deadline of I/O in QoS queue based on the responsetime requirement. Disk I/O scheduling is a strategy which supports both the requirements of storage bandwidth and responsetime.It separates metadata access from the data access in iSCSI OSD storage system. MDS (MetaData Server) manages metadata, and the amount of metadata access in storage systems takes up 50% to 80% of the entire amount of data access. So metadata service system is a potential visitting bottleneck. The performance of the entire storage system is improved by providing load balancing of metadata service, and it is also a basic guarantee for achieving responsetime QoS support. It analyzes the traditional metadata service system, and analyzes the reliability of the traditional dual MDS by theoretical analysis. Metadata mapping management based on hash twice is designed, and the metadata placed strategy is based on the idea that object metadata are stored in adjacent position in barrels. It is proved in theory to have a low overhead to find and delete. Designing a linear chain load balancing strategy in heterogeneous MDS cluster, it is proved to have a good load balancing by theoretically analyzing the comparative response time in MDS cluster and simulation test.The prototype experiment platform is builded by analysing the design principles of OSD simulation, and respectively test the cache allocation strategy with QoS support, the OSD I/O scheduling with QoS support and iSCSI OSD with end to end QoS support. Test results show that the deviation of the actual measured value with QoS support strategies and QoS expected value are less than 9%, therefore it indicates that the strategys with QoS support take good effect.