@phdthesis{12805,
  author = {Ketil Lund},
  title = {Adaptive Disk Scheduling in a Multimedia DBMS},
  abstract = {In the last years, Internet access methods such as xDSL and cable modems have become much cheaper in Europe, and the result is that broadband network access is becoming more and more widespread. Consequently, online multimedia services like News-on-Demand (NoD), digital libraries and Learning-on-Demand (LoD) are becoming important elements of the information society. The scenario of our work is a LoD-system, where a multimedia database management system (MMDBMS) realizes a large repository for multimedia-based learning material. Teachers create multimedia presentations consisting of combinations of multimedia objects, while the students search for and play back these presentations. Consequently, the MMDBMS must handle a constantly shifting workload, with very diverse requirements; and because the large size of the multimedia data implies that secondary storage must be used, these requirements are also imposed on the storage subsystem. In this thesis, we show that disk scheduling is an effective way of meeting these storage subsystem requirements. We have developed a disk scheduling framework called APEX (AdaPtive disk-schEduler for mixed-media workloads), which is specifically targeted at the requirements of a MMDBMS-based LoD-system. APEX is able to provide a number of different service types, support QoS, and at the same time achieve very high disk utilization. This has been made possible through the use of the following four techniques: (1) dynamic queue management, which keeps the overhead of the scheduling framework low, and updates bandwidth reservations according to the requirements of the MMDBMS; (2) an extended token bucket model, which ensures accurate distribution of disk bandwidth; (3) a batch building principle, which submits disk requests in batches, and ensures high disk utilization; and (4) a work-conservation feature that re-distributes unused disk bandwidth without loss of disk efficiency. We have implemented APEX together with two other disk schedulers in a simulation environment. Our measurements show that for best-effort disk requests, APEX achieves approximately 30\% higher throughput and from 30\% to 90\% lower response times than comparable schedulers, while providing the equal QoS-guarantees for real-time requests. In addition, our analysis shows that APEX does not impose any higher computational complexity than other disk schedulers offering multiple service types. From our simulations, it is also clear that APEX is able to handle the extensively shifting MMDBMS workload; it makes sure that the disk bandwidth is distributed according to the reservations that are made, and then re-distributes unused bandwidth to where it is needed. In addition, APEX is designed to take advantage of the intelligence of modern disks, leaving the final ordering of requests to the disk itself, and thereby further increasing the disk utilization. Consequently, our disk scheduling framework is well suited for a demanding environment like a MMDBMS. However, due to its modularity and dynamic configurability, APEX is also well suited for a wide range of other applications, whether they require real-time service, high-throughput, low-latency, or any combination of these.},
  year = {2003},
  month = {December},
  publisher = {Department of Informatics, University of Oslo},
}