SLEDs: Storage Latency Estimation Descriptors

Abstract

Managing the latency of storage systems is a key to creating effective very large scale information systems, such as web interfaces to satellite image databases and video-on-demand servers. Storage Latency Estimation Descriptors (SLEDs) are architecture-independent descriptions of the retrieval time of a unit of data. They describe the latency to the first byte, and the bandwidth expected. SLEDs are an important enabling technology for true end-to-end quality of service (QoS) because they can be used to predict and schedule data transfer with multimedia (guaranteed I/O rate) file systems. SLEDs provide the interface that allows database management systems (DBMS) and clients of hierarchical storage management (HSM) systems to optimize their data access patterns by choosing to read data in specific sequences or not at all. SLEDs are designed to work in intra-machine, local-area network (LAN) and wide-area network (WAN) storage systems, and to scale through twelve orders of magnitude in latency, from thousands of seconds down to nanoseconds. 1