A. L. Chervenak, D. A. Patterson, and R. H. Katz. Choosing the best storage system for video service. In Proc. ACM Multimedia'95, pages 109--119, 1995.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....device is used for caching, the performance of the MEMS cache depends on the data management policy used for the MEMS bank. With multiple devices, we must ensure that the load on the MEMS devices is balanced. In this regard, we can draw on research from data management policies for disk arrays [3]. We now investigate two cache management policies, representing two classes of load balancing strategies, which ensure total loadbalance across the MEMS bank. These approaches make different trade offs to optimize for a sub set of system configurations. More sophisticated load balancing ....

....of load balancing strategies, which ensure total loadbalance across the MEMS bank. These approaches make different trade offs to optimize for a sub set of system configurations. More sophisticated load balancing strategies, including hybrid approaches of the above, have been proposed in literature [3, 15, 23, 24]. We investigate two simple, representative approaches as a first step. 3.2.1 Striped Cache management Using striped cache management, each stream is bit or byte striped across all the k MEMS devices. There is no redundancy, and data for each stream is distributed in roundrobin fashion across ....

[Article contains additional citation context not shown here]

A. L. Chervenak and D. A. Patterson. Choosing the best storage system for video service. Proceedings of ACM Multimedia 95, pages 109--118, November 1995.

....streams that the disk array can support is limited by the the most heavily loaded disk. They develop an analytical model for the work load of the most heavily loaded disk. This model is used to determine the block size for striping that maximizes the number of supported streams. Chervenak et al. [2] study the performance of video servers in Video on Demand (VoD) systems that do not allow for any interactivity, such as VCR actions, and have a 60 second start up delay. Their study is restricted to localized placement and full striping; group striping is not considered. Chervenak et al. find ....

A. L. Chervenak, D. A. Patterson, and R. H. Katz. Choosing the best storage system for video service. In Proceedings of A CM Multimedia, 1995.

....device is used for caching, the performance of the MEMS cache depends on the data management policy used for the MEMS bank. With multiple devices, we must ensure that the load on the MEMS devices is balanced. In this regard, we can draw on research from data management policies for disk arrays [3]. We now investigate two cache management policies, representing two classes of load balancing strategies, which ensure total load balance across the MEMS bank. These approaches make di#erent trade o#s to optimize for a sub set of system configurations. More sophisticated load balancing ....

....of load balancing strategies, which ensure total load balance across the MEMS bank. These approaches make di#erent trade o#s to optimize for a sub set of system configurations. More sophisticated load balancing strategies, including hybrid approaches of the above, have been proposed in literature [3, 14, 22, 23]. We investigate two simple, representative approaches as a first step. 3.2.1 Striped Cache management Using striped cache management, each stream is bit or byte striped across all the k MEMS devices. There is no redundancy, and data for each stream is distributed in round robin fashion across ....

[Article contains additional citation context not shown here]

A. L. Chervenak and D. A. Patterson. Choosing the best storage system for video service. Proceedings of ACM Multimedia 95, pages 109--118, November 1995.

....playback stream is started, data is sequentially read from the storage system at the playout rate. This predictability is exploited in many video servers that carefully layout data on disk to achieve good load balance and high real time performance. The most widely used scheme is data stripping [1][4][15] We on other hand, have proposed a random allocation of data to disks as a mean to support more general workloads [2] 12] Nevertheless, we have shown in [12] that even for predictable workloads like video playback, despite the randomness of data allocation, our system has performance ....

A.L. Chervenak, A.A. Patterson, R.H. Katz, "Choosing the Best Storage System for Video Service", ACM Multimedia 95, p.109-119, 1995.

....networks are scalability, reliability, availability and serviceability. Availability includes the important demand that SANs can adapt quickly to a changing number of disks. Due to the fact that SANs may consist of a large number of disks, classical data placement strategies like disk striping [7] may cause severe problems. If, for instance, the striping methods de ned in the RAID levels 4 or 5 are used as a global placement strategy, any removal or insertion of a disk causes the redistribution of virtually all the data in the system. Cur 1 RAID means Redundancy Array of Independent ....

....mean time between failure (MTBF) 4] Furthermore, the used access strategy (scheduling of requests) space requirements (bu ers) and application properties (pattern of requests) have a large impact on the usefulness of such distribution strategies. The simplest data layout used is disk striping [7] which is applied with di erent granularity in a number of approaches [14, 20, 6, 8, 4] Here, the data is cut into equal sized blocks and assigned to disks in a round robin fashion so that logically consecutive blocks are put on consecutive disks, cycling repeatedly over all of them. This simple ....

A. L. Chervenak, D. A. Patterson, and R. H. Katz. Choosing the best storage system video service. In The Third ACM International Multimedia Conference and Exhibition (MULTIMEDIA '95), pages 109-120, New York, November 1996. ACM Press.

....that contention for the drives is avoided and real time guarantees can be made. There are a number of proposals for layout of video data on parallel disks. The most common method proposed is to stripe each object across the parallel disks using a fixed size stripe granule (i.e. disk block) 2] 5][8][14] 22] 27] While allocation of a disk block on a disk is often random, logically consecutive blocks are typically allocated in strictly roundrobin order to disks. This approach can work well when the workload is highly predictable, uniform and has constant bit rate (CBR) However, in practice ....

....allocation. Surprisingly, our results show that RIO is competitive and often outperforms traditional striping schemes, even for uniform sequential CBR stream workloads. 1. 2 Related Work Data striping has been proposed in many video servers, in which videos are striped over a set of disks [2] 5][8][14] 22] 27] The advantage of striping over designs in which different objects are stored on different disks is that it decouples storage allocation from bandwidth allocation, avoiding potential load imbalances due to variations in object popularity. However, striping imposes two basic ....

[Article contains additional citation context not shown here]

A. Chervenak, D. Patterson, and R. Katz. Choosing the best storage system for video service. In ACM Multimedia 95, pages 109--19, San Francisco, CA, 1995. ACM.

....that contention for the drives is avoided and real time guarantees can be made. There are a number of proposals for layout of video data on parallel disks. The most common method proposed is to stripe each object across the parallel disks using a fixed size stripe granule (i.e. disk block) 2] 4][7][13] 21] 26] While allocation of a disk block on a disk is often random, logically consecutive blocks are typically allocated in strictly roundrobin order to disks. This approach can work well when the workload is highly predictable, uniform and has constant bit rate (CBR) However, in practice ....

....allocation. Surprisingly, our results show that RIO is competitive and often outperforms traditional striping schemes, even for uniform sequential CBR stream workloads. 1. 2 Related Work Data striping has been proposed in many video servers, in which videos are striped over a set of disks [2] 4][7][13] 21] 26] The advantage of striping over designs in which different objects are stored on different disks is that it decouples storage allocation from bandwidth allocation, avoiding potential load imbalances due to variations in object popularity. However, striping imposes two basic ....

[Article contains additional citation context not shown here]

A.L. Chervenak, D.A. Patterson, R.H. Katz "Choosing the Best Storage System for Video Service", ACM Multimedia 95, San Francisco, CA, p.109-19, 1995.

....exploited by many data placement and scheduling strategies that carefully lay out data on disks in order to achieve a reasonable load balance and real time performance. See [CLOT96] for an overview and [CLOT96] WLDH96] and [VRG95] for some examples. The most wide spread scheme is data striping [BMGJ94, CPK95]. There are two basic striping techniques: either striping each data block over all disks or placing the blocks of the data space in a round robin fashion in the disk subsystems (usually RAID) of a storage network, one block per subsystem. Striping has been used, for instance, in the MARS project ....

A.L. Chervenak, A.A. Patterson, and R.H. Katz. Choosing the best storage system for video service. In Proc. of 3rd ACM Multimedia, pages 109-119, 1995.

....enormous storage requirements of modern applications, such as continuous media services, digital libraries, and scientific computing systems. Hence, research on design and evaluation of configurations and resource management techniques at various levels of the storage hierarchy is gaining momentum [23, 18, 22, 29, 1, 17]. All these works illustrate two important points: 1) it is difficult to evaluate and predict performance of storage devices and (2) the required performance characteristics are often achieved at the cost of complex solutions (be they data layout, scheduling, or other algorithms) given the ....

David A. Patterson Ann L. Chervenak and Randy H. Katz. Choosing the Best Storage System for Video Service. In Proc. of ACM SIGMETRICS Conf., pages 109--119, 1995.

....Video Service Since multimedia data are usually large, the storage system is an essential element for assuring that media stream delivery is fast and efficient. Hence an essential background in the storage devices designed for media servers is necessary. In this section, we will discuss the paper [3] which studied 3 disk arrangements for their suitability in providing VOD services. Interestingly, the disk organization suggested by the authors is used by all the video server products which we describe in Sections 3, 4 and 5 2.1 Storage Systems for Comparison In this section, we briefly talk ....

....for their suitability in providing VOD services. Interestingly, the disk organization suggested by the authors is used by all the video server products which we describe in Sections 3, 4 and 5 2. 1 Storage Systems for Comparison In this section, we briefly talk about each scheme proposed in [3] in more details. Their characteristics, advantages and weaknesses will be pinpointed. Jurassic Park Ishtar Figure 1: Video server configuration with one movie per disk. 2.1.1 One Movie Per Disk Figure 1 shows a video disk farm 1 in which one movie is stored on each disk. The authors assume ....

[Article contains additional citation context not shown here]

Ann L. Chervenak and David A. Patterson. Choosing the best storage system for video service. In Multimedia '95, pages 109--119, San Francisco, CA USA, 1995. ACM Press.

....streams that the disk array can support is limited by the the most heavily loaded disk. They develop an analytical model for the work load of the most heavily loaded disk. This model is used to determine the block size for striping that maximizes the number of supported streams. Chervenak et al. [2] study the performance of video servers in Video on Demand (VoD) systems that do not allow for any interactivity, such as VCR actions, and have a 60 second start up delay. Their study is restricted to localized placement and full striping; group striping is not considered. Chervenak et al. find ....

A. L. Chervenak, D. A. Patterson, and R. H. Katz. Choosing the best storage system for video service. In Proceedings of ACM Multimedia, 1995.

....has been studied is how to optimize the use of disk and memory. Several strategies for disk allocation has been proposed. The main strategies are either to store the entire video on a single disc, to stripe the video on several disks on a single machine [1] or on several disks on multiple machines [2, 8, 19, 30]. The main reason for introducing striping has been to achieve better load balancing on the disks. Disk scheduling and how to optimize the use of the disk bandwidth has been studied intensively [6, 32, 36] Closely related to disk allocation and disk scheduling is how to utilize the available main ....

A. L. Chervenak, D. A. Patterson, and R. H. Katz. Choosing the best storage system for video service. In Proceedings of ACM Multimedia '95, pages 109--119, San Francisco, California, November 1995.

....balanced across server clusters. 10 In Eq. 1) n Gamma 1)Tplay is the largest term. 639 Theta 0:5 = 319:5 seconds. Based on video store rental patterns, it is known that accesses to movies in the server will be highly localized, with a small number of movies receiving most of the accesses[6]. According to Zipf s Law[6] the probability of choosing the nth most popular of M movies is C=n, where C = 1= 1 1=2 1=3 Delta Delta Delta 1=M ) Thus, replicating popular movies in server clusters can keep the load of server clusters balanced. In this experiment we allocate 1,000 movies to ....

....10 In Eq. 1) n Gamma 1)Tplay is the largest term. 639 Theta 0:5 = 319:5 seconds. Based on video store rental patterns, it is known that accesses to movies in the server will be highly localized, with a small number of movies receiving most of the accesses[6] According to Zipf s Law[6] the probability of choosing the nth most popular of M movies is C=n, where C = 1= 1 1=2 1=3 Delta Delta Delta 1=M ) Thus, replicating popular movies in server clusters can keep the load of server clusters balanced. In this experiment we allocate 1,000 movies to server clusters in the ....

A. L. Chervenak, D. A. Patterson, and R. H. Katz, "Choosing the best storage system for video service," Proc. of ACM Multimedia, pp. 109--119, 1995.

....and massively parallel multiprocessors. Microsoft s Tiger [1] is based on inexpensive PCs interconnected by an ATM switch. SB PRAM [9] is an example of a massively parallel processor used as a video server for Video on Demand. Much research has been on disk allocation and disk scheduling [5, 20]. The two main strategies for disk allocation are to store the entire video on one node in the server or to distribute video on several or all of the nodes in the server. An example of the first strategy is SGI s server used in Time Warner s Orlando project [25] Microsoft s Tiger uses the other ....

A. L. Chervenak, D. A. Patterson, and R. H. Katz. Choosing the best storage system for video service. In Proceedings of the ACM Multimedia'95, pages 109--119, San Francisco, California, November 1995.

....has not received much attention. We believe that this is because one major application of multimedia servers is movies on demand where a delay of a few minutes before a new multi hour movie starts is acceptable. Indeed, some of the proposed servers either ignore the latency issue entirely [5] [9] or acknowledge having initial latencies on the order of minutes [7] However, we believe that there are important applications where high latencies are not acceptable. For example, consider a video game where at each step the player s actions determine what short video to play next. Here we ....

....our techniques. The second way is to assume that the M units represent a disk array that can be viewed as a single virtual disk with high bandwidth and striped segments. Here again, we can use our techniques to allocated the segments to the virtual disk. We refer interested readers to [1] 4] [5] for a discussion of other issues related to multiple disks. The rest of this paper is organized as follows. Section 2 describes the previously studied allocation strategies and performance model that we use as a starting point. Section 3 describes our allocation techniques for reducing initial ....

A. L. Chervenak and D. A. Patterson. Choosing the best storage system for video service. Proceedings of ACM Multimedia 95, pages 109--118, November 1995.

....and instabilities, and draw meaningful performance trends, asymptotes and limits. The model we develop is simple and allows us to address many performance and design issues of a video server of such kind. Previous work on hierarchical storage systems has been on justifying its cost advantage [54, 53, 55], or studying the possible operational procedures and their performance [56, 57, 58, 59, 60] Our work differs in addressing how the system parameters in a hierarchical storage system can be dimensioned so as to satisfy a certain user delay requirement. We have developed a simple model which can ....

....are first staged onto and then streamed from the disks. Using a cost function of storage and bandwidth, it shows that for low video popularity, video should be directly streamed from the tertiary level, and for very popular video, they should be permanently stored on disk. Chervenak et al. in [55] study three storage systems for movies on demand application: i) one movie per disk, ii) striped disk farm, and iii) hierarchical storage system. Using the metric as average cost per video stream achieving a certain criteria in the user startup delay, they find that striped disk farm performs the ....

A. L. Chervenak, D. A. Patterson, and R. H. Katz, "Choosing the best storage system for video service," in Proceedings of ACM Multimedia, pp. 109--119, 1995.

....deals with continuous media with much larger file sizes. Kienzle et al. in [23] study the cost aspect of using hierarchical storage system for video applications, and given their model they show that files of low popularity should be directly streamed from the tertiary level. Chervenak et al. in [24] use average cost per video stream to evaluate a number of storage systems. Barnett and Anido in [25] compare the cost of a centralized video system with a distributed system. Giovanni et al. consider minimizing the total cost of a hierarchical storage system in order to meet a certain user ....

A. L. Chervenak, D. A. Patterson, and R. H. Katz, "Choosing the best storage system for video service," in Proceedings of ACM Multimedia, pp. 109--119, 1995.

....been done on scheduling random accesses on digital tape. Hillyer and Silberschatz [3] presents a detailed model of DLT 4000 tapes. Based on this model they have developed and tested several scheduling algorithms [4] Using digital tape for storing digital video has been studied by several projects [2, 6, 8], but none of these have studied how to optimize random accesses to the videos stored on the tape. 2 Technologies for Digital Tape There are two main classes of digital tape technology. The first is helical scan, where the tape drive writes vertical (Figure 1a) or diagonal (Figure 1b) tracks on ....

A. L. Chervenak, D. A. Patterson, and R. H. Katz. Choosing the best storage system for video service. In Proceedings of ACM Multimedia '95, pages 109--119, San Francisco, California, November 1995.

....has not received much attention. We believe that this is because one major application of multimedia servers is movies on demand where a delay of a few minutes before a new multi hour movie starts is acceptable. Indeed, some of the proposed servers have either ignored the latency issue entirely [6, 10] or acknowledged that they involve initial latencies on the order of minutes [8] However, we believe that there are applications for which high latencies are not acceptable. For example, consider a video game in which at each step the player s actions determine what short video to play next. ....

....to the memory needed to support IO s to one disk. The total system memory requirement is thus M times the values we report. Similarly, the throughput values are M times the reported values. For a discussion of other data placement methods with M disks, interested readers can consult references [2, 4, 5, 6, 13]. The rest of this paper is organized as follows. Section 2 describes the allocation strategies and performance model that we use as a starting point. Section 3 examines some allocation and disk scheduling techniques. Section 4 proposes our on disk partial data replication scheme for reducing ....

A. L. Chervenak and D. A. Patterson. Choosing the best storage system for video service. Proceedings of ACM Multimedia 95, pages 109--118, November 1995.

....successive accesses which cycle, in round robin fashion, through all disks. The great advantage of striping is that it decouples storage capacity from disk bandwidth, since each object uses the bandwidth of all disks and thus disk overloading due to skew in object popularity is not an issue. [8] [31] 15] 33] A common approach to scheduling requests when objects are striped over all disks, is to process requests in cycles of constant duration synchronized across all disks. In each cycle, each active stream accesses a single block on a particular disk. On each cycle the data read ....

A.L. Chervenak, A.A. Patterson, R.H. Katz, "Choosing the Best Storage System for Video Service", ACM Multimedia 95, pp.109-119, 1995.

....stream is started, data is sequentially read from the storage system at the playout rate. This predictability is exploited in many video servers that carefully lay out data on disk to achieve good load balance and high real time performance. The most widely used scheme is data stripping [1][4]. We, on other hand, have proposed a random allocation of data to disks as a means to support more general workloads [2] 12] Nevertheless, we have shown in [12] that even for predictable workloads such as video playback, despite the randomness of data allocation, our system has performance ....

A.L. Chervenak, A.A. Patterson, R.H. Katz, "Choosing the Best Storage System for Video Service", ACM Multimedia 95, p.109-119, 1995.

....such that contention for the drives is avoided and real time guarantees can be made. There are a number of proposals for layout of video data on parallel disks. The most common method proposed is to stripe each object across the parallel disks using a fixed size stripe granule (i.e. disk block)[8]. While allocation of a disk block on a disk is often random, logically consecutive blocks are typically allocated in strictly round robin order to disks. This approach can work well when the workload is highly predictable, uniform and has constant bit rate (CBR) However, in practice several ....

A.L. Chervenak, D.A. Patterson, R.H. Katz "Choosing the Best Storage System for Video Service", ACM Multimedia 95, San Francisco, CA, p.109-19, 1995.

....much a server can earn by supplying video, while the latter represents the hardware cost to store a certain capacity of multimedia data. Most of the examples in this paper reflect a video server that provides movies on demand [30] 39] 35] 1] 31] 16] 32] 38] 12] 28] 14] 40] [7]. However, the results presented generalize to a large class of multimedia servers, including world wide web servers, that provide many simultaneous streams and support interactivity. The simulations assume MPEG 1 or MPEG 2 video encoding and, for simplicity, constant bit rate video. Bandwidth ....

Ann L. Chervenak, David A. Patterson, and Randy H. Katz. Choosing the Best Storage System for Video Service. In Proceedings. ACM Multimedia, November 1995.

No context found.

A. L. Chervenak, D. A. Patterson, and R. H. Katz. Choosing the best storage system for video service. In Proc. ACM Multimedia'95, pages 109--119, 1995.

No context found.

Ann L. Chervenak, David A. Patterson, Randy H. Katz. Choosing the Best Storage System for Video Service. Proceedings of ACM Multimedia '95. October, 1995.

First 50 documents

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC