A QoS degradation policy for revenue maximization in fault-tolerant multi-resolution video servers

Video servers are high@ vulnerable to disk failures because they employ large-scale disk arra.vs to support many clients and multimedia contents. To cope with such a structural weakness, they reserve a significant portion of their resource contingent on disk operability, leading to its under-utilization in normal operation mode. This paper addresses the improvement of resource utilization in faulttolerant multi-resolution video servers. For this purpose, we propose a QoS degradation method and an admission control algorithm to maximize the system performance measured as revenue while guaranteeing ever?, admitted client the minimum quality level in the event of disk failure. Our scheme exploits the multi-resolution property of video streams, and achieves graceful degradation at near-optimal level when the disk,fails. Simulation results show that, in the best case, ourproposed technique almost doubles the number ofclients admitted for video service by greatly enhancing the resource utilization rates. Index Terms Fault-Tolerance, Multi-Resolution Video Severs, Graceful Degradation, Admission Control. I . INTRODUCTION ECENT advances in multimedia and network technologies make it possible to provide video-on-demand (VOD) services to clients. Unl,ike traditional broadcasting services, clients subscribing to VOD services may control quality of service (QoS) parameters such as resolutions, frame rates and display sizes according to their own needs. Recent studies have revealed that the most practical way to satisfy such various QoS requirements of clients is to use a multiresolution video stream [6], [15], [17]. A multi-resolution or scalable video stream is a video sequence encoded in such a way that subsets of the full-resolution video stream can he decoded to extract lower-resolution streams. By using multi-resolution compression techniques, we can reduce storage requirements greatly because there is no need to keep multiple copies of the video at different resolutions in the server. We can also provide adaptable video service under the transient overloads by degrading the resolution levels. QoS degradation, however, should he implemented judiciously because there exist applications that cannot accept video streams below a certain level of resolution [4]. In addition, clients tend to request the hest possible quality R . . ' M. Song and H. Shin are with the School of Computer Science and Engineering, Seoul National University, Seoul, Korea, E-mail: [email protected], [email protected]. Contributed Paper unless the video service is coupled with monetary incentives, so several pricing mechanisms which provide monetary incentives to offset QoS degradation have been proposed [SI, [15], [16]. These pricing structures prompt video service providers to offer low-resolution videos at a low rate and charge a premium for high-resolution videos. Then clients decide the resolutions of video streams subject to their budgets and minimum quality requirements. The perceived monetary benefits to the video supplier are usually represented as revenue. Hence, revenue maximization becomes an important issue for the VOD service [SI, [15]. Generally, round-based scheduling is used for data retrieval: Time is divided into equal-sized periods, called rounds, and each admitted client is served once in each round. In addition, video servers typically employ disk arrays to support a large number of clients. In disk-array-based video servers, a video object is distributed over multiple disks to utilize disk bandwidth efficiently [21]. We will refer to this scheme as striping. The unit of data for striping, or a striping unit, denotes the maximum amount of contiguous data that will be stored on a single disk. Even though a single disk is fairly reliable; disk arrays with a large number of disks are highly vulnerable to disk failures [IO]. Thus, disk arrays usually employ reliability techniques of the redundant array of independent disks (RAID) architecture, ' which achieves fault tolerance either by replication or parity encoding [2 11. In replication-based schemes which are employed in RAID level 1, the original data is duplicated on separate disks. On the other hand, in parity-based schemes, a panty is computed by an exclusive-or operation and used for data reconstruction. We refer to the parity block together with all data blocks over which parity is computed as a parity group. The parity encoding techniques are employed in RAID levels 3, 4, and 5. In RAID 3 level, data is distributed in very small chunks across disks and a dedicated disk stores all parity information. Unlike RAID 3 level, RAID level 4 interleaves data in fixed-size blocks with a dedicated disk storing parity information. In RAID 5 level, data is distributed in'fixed-size block, and parity blocks are distributed across all disks. Generally, there exist two service modes of disk arrays [IO]: (I) normal mode, where all disks are operational, and (2) degraded mode, where a disk has failed. Fault-tolerant design of video servers is complicated for the following two reasons: First, since the server cannot access data on a failed disk, additional server resources are required for the reconstruction of data. Second, the retrieval of video objects should be performed in real-time even in the presence Manuscript reccived April 15,2003 0098 3063100 $10.00 D 2003 IEEE M. Song and H. Shin: A QoS Degradation Policy for Revenuc Maximization in Fault-Tolerant Multi-Resolution Video Servers 393 of a disk failure. To resolve these problems, current approaches reserve contingent bandwidth or huffer [3], [9], [lo], [19]. As a consequence, they under-utilize the server resources in normal mode. In this paper, we propose a QoS degradation method for multi-resolution video servers, which handles revenue maximization and fault-tolerance simultaneously. Instead of wasting server resources for the degraded mode, our scheme exploits the multi-resolution property of streams and provides QoS-negotiated service in the face of disk failure. We first develop a QoS model which defines a spectrum of resolutions for video streams and permits quality degradation when a disk fails; thus, it can improve resource utilization in normal mode. For graceful degradation, we address the resolution degradation problem in order to maximize overall system performance, measured as revenue. In other words, a revenue is defined by what the client is willing to pay given the quality of the video received; the quality degradation is done in a way so as to maximize the revenue accrued. We prove that the problem is NP-hard and propose heuristic algorithms to find resolutions of accepted streams in degraded mode. An admission control algorithm is also presented to guarantee that every client admitted can he serviced to some per-client specified minimum quality level even in degraded mode. The rest of this paper is organized as follows: In Section 11, we briefly explain related work. In Section 111, we present system models for further description of our schemes. Next, we propose an admission control algorithm and resolution adjustment algorithms in Section IV and V. We validate the proposed schemes through simulations in Section VI. We extend our scheme for variable-hit-rate (VBR) streams in Section VII. Finally we conclude the paper in Section VIII.