تجزیه و تحلیل عملکرد از سرور ذخیره سازی ویدئو تحت مرزهای تاخیر اولیه
|کد مقاله||سال انتشار||مقاله انگلیسی||ترجمه فارسی||تعداد کلمات|
|27414||2000||17 صفحه PDF||سفارش دهید||7189 کلمه|
Publisher : Elsevier - Science Direct (الزویر - ساینس دایرکت)
Journal : Journal of Systems Architecture, Volume 46, Issue 2, 15 January 2000, Pages 163–179
Previous studies on video storage servers focused on improving the disk throughput and reducing the server buffer size. However, the initial delay of a new request, one of the most important quality of service (QoS) parameters from the users' point of view, is almost neglected while designing a storage subsystem or evaluating its performance. For different types of video-on-demand (VOD) services such as interactive video game, digital library, or movie-on-demand system, the initial delay can vary from 0.5 s to 5 min. This criterion brings some impacts on designing a storage server for a particular VOD application. In this paper, we investigate the storage server design and the performance evaluation of VOD systems with different initial delay guarantees. We propose a new performance model on evaluating the efficiency of a video storage server so that a cost-effective configuration can be easily obtained under a specified initial delay bound.
Recently, the video-on-demand (VOD) system has been widely applied to entertainment and education. The most important design issues of a video storage server are to provide jitter-free video services as well as to promote the utilization of the storage bandwidth to accommodate more users. Disk array system which stripes video files on several hard disks is a common approach to increase the space and bandwidth capacities and to support a large number of users accessing a huge amount of video content for a video storage server 12, 14 and 21. Two basic data striping techniques, fine-grain and coarse-grain, have been employed on a disk array. For fine-grain striping, a video file is divided into access blocks, and each block is further striped onto a number of hard disks. Here, an access block is defined as the total amount of data retrieved once by a storage server for a particular read request. Therefore, the hard disks can serve a single read request parallel to their bandwidths. On the other hand, for coarse-grain striping, each access block is completely stored on a hard disk in order to increase the number of concurrent users served by the storage server at a time. In general, coarse-grain striping has a higher concurrency but a poorer parallelism than fine-grain striping. Oezden et al. studied on data striping techniques for VOD applications and concluded that coarse-grain striping is more suitable for the VOD system than fine-grain striping due to its low buffer requirement and high disk throughput . To improve the disk throughput, disk scheduling policies and data placement schemes are widely exploited. Round-robin disk scheduling services periodical requests by a fixed order without taking the physical location of these requested blocks into account. This policy obtains fast response time, low buffer requirement, but results in poor disk bandwidth utilization owing to non-optimized disk seeking 7, 11 and 18. Different from round-robin scheduling, SCAN disk scheduling optimizes the block retrieval sequence by their physical locations. Thus, it reduces seek time overhead and promotes the disk throughput to a great extent. However, more delay will be introduced for a request, and since the retrieval order of the requested blocks is variant by the time, each user requires one additional buffer to compensate the uncertainty of data arrival time into the memory buffer . Yu et al. proposed grouped sweeping scheduling (GSS) to get the compromise between the above two policies . They partition access requests into several groups, apply round-robin between the groups and employ SCAN within a group. GSS has a higher disk throughput than a round-robin, and a lower buffer requirement and delay than SCAN. Occasionally, data placement schemes are employed to promote disk throughput, which can be categorized into three approaches: random, contiguous, and constrained data placement schemes. Contiguous data placement scheme sequentially stores a video file on a hard disk in order to eliminate extra seeks while retrieving an access block, which obtains a much higher disk throughput than the random data placement scheme. Rangan and Vin studied on the fundamental disk layout problem for a multimedia application, and presented a constrained data placement method which places consecutive blocks of the same media file within a fixed distance to avoid jitters on playback . Some variants of constrained data placement called region-based or cluster-based schemes were explored in Refs. 3, 4 and 5. They partition a disk into a number of regions where a region is defined as a group of contiguous tracks on a disk, and store the blocks of a video file on these regions by a pre-determined sequence. Then, the disk head is confined to retrieve data blocks within a region during a period of time. After the period of time, the disk head moves to the nearby region and reads the requested blocks for the next period. By restraining the accesses within a region of a hard disk, the disk throughput can be improved by reducing the seeking distance. However, the users suffer from long initial delays. Change and Hector Garcia-Molina proposed several solutions on reducing initial delay for region-based data placement, but they require extra storage space to replicate the first several blocks of each video file . To evaluate the performance of a video storage server, some work emphasized the trade-off between data striping strategies, buffer requirement, and disk throughput . Some studies presented performance models based on the number of supported users per hard disk or cost-effective models based on cost per user 9 and 10. Unfortunately, the initial delay of new requests is never taken into consideration among them. Not only the jitter-free during playback but also the initial delay, one of the quality of service (QoS) parameter, are most concerned by users for the demand service. The acceptable initial delay varies from less than 0.5 s to more than 5 min for different applications. For example, two video clips should be smoothly connected for interactive video game system, and more than half seconds delay between two video clips may introduce sensible visual distortion to game players . For a digital library system, the delay of a request probably should be less than 3 s for on-line users. For the movie-on-demand system or near VOD system, longer delays can be allowed, since 3–5 min of delay is reasonable for viewing an around 120 min movie. In the previous studies on video storage server design, however, they either promoted disk throughput or reduced delays for a particular application and a layout scheme. The work on the disk configuration of a video storage server based on the given initial delay bounds still lacks study yet. In this paper, we investigate the performance model, evaluation, and analysis of a video storage server for VOD systems with different initial delay requirements. We propose a new performance model for a disk based storage server, where striping strategies, buffer requirement, disk throughput, and initial delay are all taken into account at the same time. By applying the proposed model, the cost-effective storage configuration for a particular VOD system can be easily obtained. The rest of the paper is organized as follows. Section 2describes the storage subsystem for a video server. In the section, we also introduce the underlying framework of some disk scheduling and layout schemes. Section 3presents the proposed performance models and Section 4illustrates and discusses the simulation results. Finally, we conclude this study in Section 5.
نتیجه گیری انگلیسی
In this paper, we proposed a new performance model for a video storage server under specified initial delay bounds which are variant for different VOD applications. Our model takes striping strategies, disk scheduling policies, data placement schemes, block sizes, buffer requirement per user and initial delay time into account simultaneously. In that way, a system designer can easily obtain the efficient configuration of a video storage server by applying our proposed model based on the given system parameters. According to simulation results, we remark some summaries. 1. Applying round-robin scheduling on region-based data placement can achieve 20% performance improvement over applying round-robin scheduling alone, but the improvement becomes insignificant while the number of regions is more than 8. The 20% performance improvement is based on the hard disk model adopted in our experiments, and it will be decreased under the hard disk drives with higher bandwidth and less seek time overhead. 2. Applying SCAN scheduling on region-based data placement seems not practical because it improves merely 4% in number of supported users more than applying SCAN alone, but introduces more delay. 3. In general, SCAN scheduling has a better performance on disk throughput than round-robin scheduling. However, for region-based data placement with more than 4 regions, round-robin scheduling achieves a higher number of supported users than SCAN on the same buffer requirement. 4. Coarse-grain striping has the best performance on disk throughput but suffers from long delay while fine-grain striping has fast response time but results in poor disk throughput. The combination of the above two strategies is somehow a good configuration for tight initial delay VOD systems. 5. For a long initial delay or a non-guaranteed initial delay VOD system, the disk throughput is the major dominant factor on the maximal number of supported users by a video storage server. Coarse-grain striping strategy is the best solution in terms of the number of supported users and the buffer requirement.