J.L. Wolf, P. S. Yu and D. Shachnai, \DASD Dancing: A Disk Load Balancing Optimization Scheme for Video-on-Demand Computer Systems," SIGMETRICS'95, pp. 157-166, 1995. 29  Home/Search   Document Details and Download   Summary   Related Articles   Check

....component that monitors and controls available resources. When resources (memory and disk bandwidth) are exhausted, it queues newly arrived requests in order to prevent hiccups. A number of studies have outlined the design and implementation of such a component [GM98, GZS 97, YCK93, OBRS94, WSY95, TPBG93] This section begins with and how to derive its optimal value in order to minimize system cost. Next, we present BMDT, followed by a discussion of the configuration planner. For each, we describe the impact of the VCR operations. The configuration planner and BMDT are designed to ....

....stream for a clip lags another stream for the same clip by only a short time interval, then the system could retain the portion of the clip between the two in buffers. The lagging stream would read from the buffers and not have to read from disk. batching of requests [DSST94, DSS94b, OBRS94, WSY95] in this method, requests are delayed until they can be merged with other requests for the same clip. The merged streams then form one physical stream from the disk and consume only one set of buffers. Only on the network will the stream split at some point for delivery to the individual display ....

J. Wolf, H. Shachnai, and P. Yu. DASD Dancing: A Disk Load Balancing Optimization Scheme for Video-on-Demand Computer Systems. In Proceedings of the ACM SIGMETRICS and Performance, May 1995.

....are multiple ways of configuring the hardware and organizing data. These choices are a tradeoff in MTTSL, cost, and need for detective techniques that dissolve bottlenecks by replicating the data. In this study we investigate three alternative organizations: 1. Independent subservers [2] 7] [10]: With this organiza tion, a heterogeneous collection of disks is organized into a collection of subservers, each consisting of a homogeneous array of disk drives. A file (e.g. a movie) is assigned to one subserver, see Figure l(a) Hot read only files (e.g. popular movies) might be replicated ....

J.L. Wolf, P.S. Yu, and H. Shachnai, "DASD Dancing: A Disk Load Balancing Optimization Scheme for Video-on-Demand Computer Systems," in Proceedings of the A CM SIGMETRICS, Ottawa, Canada, May 1995.

....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 ....

....streams on a k device MEMS bank is similar to scheduling streams on a disk array. However, using a MEMS bank as a disk buffer must consider the real time requirements between the disk and MEMS storage as well as between MEMS storage and DRAM. On the other hand, disk load balancing policies [3, 15, 23, 24] are directly applicable to a MEMS bank when it is used as a cache. We investigated two representative policies from previous work on disk arrays for a MEMS bank. 7 Conclusions and Future Work We have investigated the possibility of using MEMS storage for buffering and caching streaming ....

J. L. Wolf, P. S. Yu, and H. Shachnai. DASD Dancing: A disk load balancing optimization scheme for video-ondemand computer systems. Proceedings of the 1995.

....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 ....

....streams on a k device MEMS bank is similar to scheduling streams on a disk array. However, using a MEMS bank as a disk bu#er must consider the real time requirements between the disk and MEMS storage as well as between MEMS storage and DRAM. On the other hand, disk load balancing policies [3, 14, 22, 23] are directly applicable to a MEMS bank when it is used as a cache. We investigated two representative policies from previous work on disk arrays for a MEMS bank. 7 Conclusions and Future Work We have investigated the possibility of using MEMS storage for bu#ering and caching streaming ....

J. L. Wolf, P. S. Yu, and H. Shachnai. DASD Dancing: A disk load balancing optimization scheme for videoon -demand computer systems. Proceedings of the 1995.

....RePacking policy described in this paper. The benefit of this additional storage cost is increased scalability and graceful degrada tion of service when nodes fail. In the following section, we describe some existing dynamic replication policies. 2. 1 Related Work The DASD Dancing scheme [13] performs load balancing by determining the most suitable node to supply a new stream and reassigning existing streams to different nodes when necessary. Related to this scheme, and of most relevance to our research, is the scheme used to assign multimedia files to nodes. The assignment begins by ....

....load imbalance. 3.3 Service assignment As client requests arrive, they are simply assigned to the least utilised node with a copy of the requested file. An enhancement of the policy might attempt to move existing streams from one node to another, in a similar way to the DASD Dancing policy [13]. M o Mo S O S S 2 (a) Initial state from previous pack M21M31M I S O S S2 (b) Repack M o M lM71 ,I1 S S O S (e) Repack M 4 (f) Re order nodes by shortfall Repack M e (c) Repack M 2 M on S o until the shortfall is less than S S S O (g) Repack M s then M7, ....

J. L. Wolf, P.S. Yu, and H. Shachnai. DASD dancing: A disk load balancing optimization scheme for video-on-demand computer systems. In Proceedings of ACM SIGMETRICS '95, pages 15166, May 1995.

....is in also considered in homogeneous configurations, when object copies are completely stored in just one or in a subset of the disks. In this case, disks storing very popular or hot objects may be overloaded and additional copies of these objects are made on other disks, to improve load balance [16]. In [5] the authors address the heterogeneity problem by placing a higher percentage of hot data on disks with higher BSR than on disks with lower BSR, such that the predicted load on disks are made proportional to their bandwidths. The drawback of this approach is that predicting object ....

J.L. Wolf, P.S. Yu, H. Shachnai, "DASD Dancing: A Disk Load Balancing Optimization Scheme for Video-on-Demand Computer SYSTEMS". SIGMETRICS

.... videos on the different resources, replicating possibly videos [Little Venkatesh 1993] The request placement policy, which is implemented by the Resource Manager, consists in dynamically assigning client requests to resources according to the requested video and to load balancing consideration [Wolf et al. 1994]. In the context of a video on demand service, several client requests for the same video may arrive at the service in a small time interval. The service can here group these requests, making them handled as a single request by the server, which multicasts the video stream to the corresponding set ....

J. Wolf, P. Yu, and H. Shachnai. DASD dancing: a disk load balancing optimization scheme for video-on-demand computer systems. Technical report. IBM Research Division. T.J. Watson Research Center. 1994.

....of the service provider, and the other called data delivery manages the delivery of video data to users. The request admission is made up of the Request manager, which batches the clients requests [AWY96] and the Resource manager which places the requests on the server s resources [SGB96, WYS94] The above constituents of a VoD service can interact using several communication links. For example, communication between the STB and the service provider can be established through a phone line or a network. The server and the service provider can be interconnected by a dedicated network and ....

....is generally daily updated in order to tackle the variation of requesting probabilities. ffl The request placement policy, which consists in dynamically assigning physical servers to the resources managing the different disks storing the requested video, according to load balancing consideration [WYS94] Irisa Providing Reliability on Demand in Distributed VoD Servers 13 Our resource manager has to consider both the reliability requirements and the accessibility improvement. In fact, this particular has no noticeable influence on the video placement policy. On the other hand, the request ....

[Article contains additional citation context not shown here]

J.L. Wolf, P.S. Yu, and H. Shachnai. DASD dancing: a disk load balancing optimization scheme for Video-on-Demand computer systems. Technical report, IBM research division, T. J. Watson Research Center, December 1994. PI n 1139

....structures and meta architectures [2] and runtime systems such as Broadway [33] and the Actor Foundry. Multimedia QoS enforcement has been a topic of extensive research. Related work in this area includes projects such as Omega [27] QualMan [26] and several algorithms for MM server management [45, 42, 22, 16, 11, 43, 10]. QualMan [26] is a QoS aware resource management platform which contains a set of resource brokers that provides negotiation, admission and reservation capabilities for sharing end system resources such as CPU, memory and network bandwidth. Much of the work on formal models for QoS has been in ....

J. L. Wolf, P. S. Yu, and H. Shachnai. Dasd dancing: A disk load balancing optimization scheme for video-on-demandcomputer systems. In Proceedings of ACM SIGMETRICS '95, Performance Evaluation Review, pages 157--166, May 1995.

.... buffer management policies to minimize memory requirements [66, 104] replication and striping strategies for optimizing storage across disk arrays [90, 151] batching mechanisms that group closely spaced requests for the same objects [55] load balancing mechanisms for effective utilization [170, 163, 57]. 2.5 Internet Computing (Metacomputing) Various systems use resources on the Internet for wide area parallel computing, e.g. Wax, Legion, Atlas [20] Globus [31] The Legion system [71] provides a unified view of a global network of computers using an extensible object model and services such ....

....7.1. 162 Load balancing and resource management have been studied mainly in the context of switched storage video servers that comprise of processing unit(s) that access a set of shared storage devices. Considering the storage subsystem alone, a two stage DASD dancing scheme for load balancing [170] uses an initial static stage followed by a dynamic phase. In the initial static stage, a greedy assignment of videos to disk groups is obtained using a graph theoretic approach. The dynamic phase that follows uses the static assignment to perform real time disk scheduling effectively. The goal of ....

[Article contains additional citation context not shown here]

Joel L. Wolf, Philip S. Yu, and Hadas Shachnai. Dasd dancing: A disk load balancing optimization scheme for video-on-demand computer systems. In Proceedings of ACM SIGMETRICS '95, Performance Evaluation Review, pages 157--166, May 1995.

....broadband environments in which hundreds of thousands of residential subscribers connect to the network to access video information, scalability of the video servers is likely to become a predominant issue. Related work in load balancing and resource management for video servers can be found in [16, 9]. The DASD dancing scheme proposed in [16] performs disk load balancing in two stages. An initial static stage determines assignment of videos to groups of shared disks. The process consists of building a clique tree whose root nodes contain the most popular videos. A greedy approach tries to ....

....thousands of residential subscribers connect to the network to access video information, scalability of the video servers is likely to become a predominant issue. Related work in load balancing and resource management for video servers can be found in [16, 9] The DASD dancing scheme proposed in [16] performs disk load balancing in two stages. An initial static stage determines assignment of videos to groups of shared disks. The process consists of building a clique tree whose root nodes contain the most popular videos. A greedy approach tries to maximize distance b w copies of the same ....

Joel L. Wolf, Philip S. Yu, and Hadas Shachnai. Dasd dancing: A disk load balancing optimization scheme for video-on-demand computer systems. In Proceedings of ACM SIGMETRICS '95, Performance Evaluation Review, pages 157--166, May 1995.

....value for distance threshold to minimize the system cost. Experimental results are presented in Section 5. The conclusions and future research directions can be found in Section 6. 2 RELATED WORK Except buffer sharing, there are two other approaches to overcome the I O bottleneck: batching [5, 16]: in this method, requests are delayed until they can be merged with other requests for the same video. These merged streams then form one physical stream from the disk and consume only one set of buffers. Only on the network will the streams split at some point for delivery to the individual ....

J. Wolf, H. Shachnai, and P. Yu. DASD Dancing: A Disk Load Balancing Optimization Scheme for Video-on-Demand Computer Systems. In Proceedings of the ACM SIGMETRICS and Performance, May 1995.

.... each video across all the nodes in the system and thus avoid the problem of splitting resources , e.g. as in the staggered striping technique [1] However, this approach suffers from a number of implementation related shortcomings that are detailed in [4] An alternate system is described in [14] where the nodes are connected in a shared nothing manner [13] Each node j has a finite storage capacity, C j (in units of continuous media (CM) objects) as well as a finite load capacity, L j (in units of CM access streams) These nodes are constructed by combining groups of disks into disk ....

....to support requests for a CM object is a function of the demand. This should result in a scalable system which can grow on a node by node basis. The difficulty here is in deciding on: 1) how many copies of each video to keep, which can be determined by the demand for that video, e.g. as in [14], and (2) how to place the videos on the nodes so as to satisfy the total anticipated demand for each video within the constraints of the given storage system architecture. Our data placement problem tries to capture these issues. 1.4 Organization We start with an overview of the sliding window ....

J. Wolf, H. Shachnai, and P. Yu. DASD Dancing: A Disk Load Balancing Optimization Scheme for Videoon -Demand Computer Systems. In ACM SIGMETRICS /Performance Conf., pages 157--166, 1995.

....sufficient bandwidth capacity to service the demand for these objects. Of course, a disadvantage of this approach is a need for additional storage space. Furthermore, techniques are needed for adjusting the number of replicas as the access patterns change. Some of these issues are addressed in [21], in the context of workloads with relatively infrequent changes in object access patterns as well as in our previous work [13, 5] where we propose dynamic replication techniques in the context of more frequent changes in data access patterns. In this paper, we improve on our previous work on ....

....corresponding existing successful implementations employ only tens of disks. In contrast, the use of replication for the purpose of addressing workload demand problems has been less explored. In [19] the authors consider skews in data access patterns but in the context of a static environment. In [21], the authors address various questions arising in the context of load imbalance problems due to skews in data access patterns, but in a less dynamic environment (than we investigate here) We believe that the policies used in this paper can be complementary to the techniques developed in [21] In ....

[Article contains additional citation context not shown here]

J. Wolf, H. Shachnai, and P. Yu. DASD Dancing: A Disk Load Balancing Optimization Scheme for Video-on-Demand Computer Systems. In ACM SIGMETRICS/Performance Conf., 1995.

....balancing and resource management have been studied mainly in the context of switchedstorage video servers that comprise of processing unit(s) with shared access to a set of storage devices. Considering the storage subsystem alone, a two stage DASD dancing scheme for load balancing is studied in [16]. In the initial static stage, a greedy assignment of videos to disk groups is obtained using a graphtheoretic approach. The dynamic phase that follows uses the static assignment to perform real time disk scheduling effectively. The goal of the dynamic phase is the minimization of an objective ....

.... and network interface(s) Since the data sources are connected via an external network and may have heterogeneous and distinct resources that are not shared with the other data sources, the tradeoffs to be considered for load management differ from those that exist in switched storage servers[16, 7]. For instance, the application of the dynamic segment replication scheme in the context of a distributed video server requires renegotiation and reestablishment of network connections for every request migration, thereby introducing unacceptable jitter during playback. We define and formulate ....

J. L. Wolf Wolf, P. S. Yu Yu, and H. Shachnai. Dasd dancing: A disk load balancing optimization scheme for video-on-demand computer systems. In Proceedings of ACM SIGMETRICS '95, Performance Evaluation Review, pages 157--166, May 1995.

....the placement of the videos on the different disks and possibly in managing the replication of videos. The request placement policy consists in dynamically assigning client requests to the CPUs managing the different disks storing the requested video, according to load balancing consideration [Wolf 94] This policy is implemented by a resource manager (see figure 1) Existing video placement policies rely on the replication of different video streams according to their requesting probabilities [Keeton 93, Little 93] More precisely, the number of copies of each video stream is an increasing ....

....to evaluate the current load of the resource. During the placement of a primary process, the allocation always takes one slot (i.e. primary : primary 1) The cost of this placement is therefore evaluated by taking into account the relative accessibility of each video based on the work of [Wolf 94] Briefly stated, in order to maximize the relative accessibility of each video, the available resource manager balances the load between the resources by maintaining as close as possible the accessibility ratios between the current load (i.e. primary secondary) and the maximal load (i.e. ....

J.L. Wolf, P.S. Yu, and H. Shachnai. DASD dancing: a disk load balancing optimization scheme for Video-on-Demand computer systems. Technical report, IBM research division, T. J. Watson Research Center, December 1994.

....the load imbalance across partitions, consider a disk array consisting of D disks that is partitioned into groups of d disks each. Let us assume that the server employs a placement policy that assigns streams to partitions such that each partition is equally likely to be accessed by a new request [8, 23]. That is, the probability that a newly arriving client accesses a partition is q = d=D. In such a scenario, if n clients access the server, then the probability that m clients access the j th partition is binomially distributed, and is given as: P (Y j = m) n m Delta q m Delta (1 ....

....the VRG model. replicating media streams across array partitions on the response time. The problem of determining the partition size was not addressed in the paper. The problem of assigning media streams to array partitions subject so as to balance the load across partitions has been dealt in [8, 23]. These efforts complement our work since they do not deal with the issue of determining an optimal partition size for large disk arrays. Many other striping related issues that are complementary to the problem addressed in this paper have been investigated. Striping techniques that minimize ....

J. Wolf, P. S. Yu, and H. Shachnai. DASD Dancing- A Disk Load Balancing Optimization Scheme for Video-on-Demand Computer Systems. In Proceedings of ACM SIGMETRICS'95, pages 157--166, 1995.

....static or dynamic. The static schemes assume that the access frequency for each media stream is known a priori and employ a placement algorithm that allocates media streams to partitions (without replication) so as to ensure that each partition is equally likely to be accessed by a new request [8, 20]. The dynamic schemes, on the other hand, enhance the static schemes by: 1) maintaining multiple replicas of media streams across partitions, and (2) servicing a new request using a replica stored on an under loaded partition, thereby balancing the load across partitions [20] Static schemes are ....

.... by a new request [8, 20] The dynamic schemes, on the other hand, enhance the static schemes by: 1) maintaining multiple replicas of media streams across partitions, and (2) servicing a new request using a replica stored on an under loaded partition, thereby balancing the load across partitions [20]. Static schemes are simple to implement, but can yield imbalance across partitions. In contrast, dynamic schemes achieve equitable load distribution across partitions, albeit at the expense of larger storage space requirement and more complex storage space management algorithms. Due to the ....

[Article contains additional citation context not shown here]

J. Wolf, P. S. Yu, and H. Shachnai. DASD Dancing- A Disk Load Balancing Optimization Scheme for Video-onDemand Computer Systems. In Proceedings of ACM SIGMETRICS'95, pages 157--166, 1995.

....for future work. Firstly, we are extending the model to take into account the restoration of availability properties after a failure. Secondly, we are extending the proposed customization for the VoD server to cope with the issue of scaling; this requires to solve the problem of load balancing [17] in order to maximize the number of accepted requests. ....

J. Wolf, P. Yu, and H. Shachnai. DASD dancing: a disk load balancing optimization scheme for Video-on-Demand computer systems. Technical report, IBM research division, Dec. 1994.

....All theoretical results in this paper are stated without proof due to space constraints. The details can be found in the full version of this paper [16] 2 Related Work Resource scheduling issues in CM databases have attracted considerable interest from the research community in recent years [10, 11, 17, 22, 25, 27, 29, 34]. However, little attention has been paid in the multimedia literature to the on line nature of the admission control problem for CM database servers. Long and Thakur [23] present simple adversary arguments to show that no on line algorithm can achieve a constant competitive ratio in the context ....

.... can be mapped to more than one servers (i.e. replication of clips is allowed) Traditionally, the goal of data placement schemes in this setting is to balance the expected bandwidth load (according to the popularities fp i g) across the available servers under the given server storage constraints [11, 22, 34]. This model of popularity based data placement aims at achieving good system utilization and balanced system load in an average sense. On the other hand, our competitiveness results indicate that to ensure robust system performance, a placement strategy should also try to achieve some sec7 0 ....

[Article contains additional citation context not shown here]

Joel L. Wolf, Philip S. Yu, and Hadas Shachnai. "DASD Dancing: A Disk Load Balancing Optimization Scheme for Video-on-Demand Computer Systems". In Proceedings of the 1995 ACM SIGMETRICS Conference on Measurement & Modeling of Computer Systems, Ottawa, Canada, May 1995.

....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 during the previous ....

J.L. Wolf, P.S. Yu, H. Shachnai, "DASD Dancing: A Disk Load Balancing Optimization Scheme for Video-on-Demand Computer SYSTEMS". SIGMETRICS 1995,pp.157-166, 1995.

....and need for detective techniques that dissolve bottlenecks by replicating the read only data. It is beyond the focus of this workshop paper to investigate all possible organizations. Instead, we provide a preliminary study of three alternative organizations: 1. Independent subservers [GS93, DS95, WYS95] With this organization, a heterogeneous collection of disks is organized into a collection of subservers, each consisting of a homogeneous array of disk drives. A file (e.g. a movie) is assigned to one subserver, see Figure 1(a) Hot read only files (e.g. popular movies) might be replicated ....

J.L. Wolf, P.S. Yu, and H. Shachnai. DASD Dancing: A Disk Load Balancing Optimization Scheme for Video-on-Demand Computer Systems. In Proceedings of the ACM SIGMETRICS, Ottawa, Canada, May 1995.

....sequentially, will generate accesses cycling 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. 6] 24] 11] [25] (Note that RIO has the same property) There are basically two approaches to achieve load balancing on striped systems. One possibility is to use large reads requesting data from all disks at a single request for each active stream [24] This guarantees that the disks will all have the same load, ....

J.L. Wolf, P.S. Yu, H. Shachnai, "DASD Dancing: A Disk Load Balancing Optimization Scheme for Video-on-Demand Computer SYSTEMS". SIGMETRICS 1995,pp.157-166, 1995.

....the requests are made for viewing a small number of hot movies. These factors, together with the importance of fault tolerance in real time VOD applications, suggest that replicating certain frequently accessed movies in some disk arrays is a viable approach to providing the VOD service required [2, 3, 6, 17]. It is worth mentioning that data replication in VOD system is employed primarily to support high I O bandwidth required for multimedia data. This is in contrast to data replication in OLTP (On Line Transaction Processing) systems that is mainly used to provide fault tolerance. The ....

J.L. Wolf, P.S. Yu, and H. Shachnai. Dasd dancing: A disk load balancing optimization scheme for videoon -demand computer systems. Proceedings of ACM SIGMETRICS, pages 157--166, May 1995.

No context found.

J.L. Wolf, P. S. Yu and D. Shachnai, \DASD Dancing: A Disk Load Balancing Optimization Scheme for Video-on-Demand Computer Systems," SIGMETRICS'95, pp. 157-166, 1995. 29

No context found.

Joel L. Wolf, Philip S. Yu, Hadas Shachnai, \DASD Dancing: A Disk Load Balancing Optimization Scheme for Video-on-Demand Computer," ACM SIGMETRICS Conference, pp. 157-166, 1995.

No context found.

Joel L. Wolf, Philip S. Yu, and Hadas Shachnai. DASD dancing: a disk load balancing optimization scheme for video-on-demand computer systems. In Proceedings of the ACM International Conference on Measurement and Modeling of Computer Systems (SIGMETRICS), pages 157--166, Ottawa, Ontario, Canada, 1995.

No context found.

J. Wolf, P. Yu, and H. Shachnai. DASD dancing: A disk load balancing optimization scheme for video-on-demand computer systems. In Proceedings of the 1995.

No context found.

J. Wolf, H. Shachnai and P. Yu, "DASD Dancing A Disk Load Balancing Optimization Scheme for Videoon -Demand Computer Systems", Proceedings of the ACM SIGMETRICS and Performance Conf., May 1995.

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