H. M. Vin, P. J. Shenoy, and S. Rao. Efficient Failure Recovery in Multi-Disk Multimedia Servers. In Proceedings of the 25th International Symposium on Fault Tolerant Computing Systems, Pasadena, CA, pages 12--21, June 1995.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....results computed from the derivations and analyzes the sensitivity of the proposed algorithms and protocols to several key system parameters. Finally, Section XI concludes the paper. II. RELATED WORKS While many studies have investigated disk level fault tolerance in video servers (e.g. 19] [22]) only three studies [5] 15] 16] known to the author have investigated server level fault tolerance in parallel video servers. In this section, we briefly review these studies and compare them with the approach proposed in this paper. Bolosky et al. 5] proposed the use of data mirroring to ....

H. M. Vin, P. J. Shenoy, and S. S. Rao, "Efficient failure recovery in multi-disk multimedia servers," in Proc. 25th Ann. Symp. Fault Tolerant Computing (FTCS'95), Pasadena, CA, pp. 12--21.

....servers studied by Berson, Golubchik and Muntz [3] employ a RAID 5 data layout and utilize a modified and expanded read schedule in degraded mode requiring additional buffering. There can be transition difficulties in which data is not delivered on time. A method proposed by Vin, Shenoy and Rao [28] does not rely on RAID 5 parity redundancy but on redundancy properties of the video data itself. Streaming RAID of Tobagi, Pang, Baird, and Gang is one of the first commercial video servers [26] There has been considerable work on disk array declustering as well; much of this centers on studies ....

Vin, H.M., P.J. Shenoy, and S. Rao. Efficient Failure Recovery in Multi-disk Multimedia Servers. In Proceedings of the 25th International Symposium on Fault-Tolerant Computing, pp. 12-21, 1995. 18

....yield service times (i.e. the total time spent in retrieving images during a round) that exceed T , resulting in playback discontinuities at client sites. Alternatively, consider a multimedia server that always computes parity over a sequence of (G 1) media blocks from the same video stream [32]. That is, all (G 1) media blocks within a parity group are successive blocks of the same video stream. In such a scenario, the server can exploit sequentiality of media playback to minimize the increase in load due to a failure by using media blocks retrieved during playback for failure ....

Vin HM, Shenoy PJ, Rao S (1995) Efficient Failure Recovery in Multi-Disk Multimedia Servers. In: Proceedings of the 25th International Symposium on Fault-Tolerant Computing Systems, June 1995, Pasadena, Calif., IEEE Press, pp 12--21

.... Determining the Degree of Image Partitioning It has been shown that a necessary and sufficient condition to ensure that none of the blocks contained in a sub image are in the 8 neighborhood of each other, the image must Storage Architectures for Digital Imagery 9 be partitioned into 4 sub images [44]. Notice that when an image is partitioned into 4 sub images, each sub image contains 25 of the image data in the frequency domain. Consequently, if the information contained in a sub image is not available, the image will have to be reconstructed from the remaining 75 of the data. Since the ....

H. M. Vin, P. J. Shenoy, and S. Rao. Efficient Failure Recovery in Multi-Disk Multimedia Servers. In Proceedings of the 25th International Symposium on Fault Tolerant Computing Systems, Pasadena, CA, pages 12--21, June 1995.

....disadvantage of this scheme is that there is a transition phase as the parity group switches to degraded mode; during this phase, some of the data is not delivered and consequently hiccups occur. The scheme that we propose (SID) provides a smooth transition to degraded mode. Vin, Shenoy and Rao [20] proposed a method which does not rely on errorcorrecting codes such as parity, but on properties of the video data itself. Their technique is a combination of a compression algorithm and a disk array layout. The 52 compression algorithm partitions the data in such a way that an acceptable (but ....

Vin, H.M., P.J. Shenoy, and S. Rao. Efficient Failure Recovery in Multi-disk Multimedia Servers. In Proc. 25th Int'l Symp. on Fault-Tolerant Computing, pp. 12--21, 1995.

.... Fundamentally, both techniques achieve fault resilience by reducing the resources required by a video stream (measured by its bandwidth) Content alteration techniques take advantage of the intra image redundancy whereas frame dropping techniques take advantage of inter image redundancy [9, 10, 12]. An alternate approach is to provide fault resilience by building redundancy into the system as in RAID based servers [2, 10, 9] Rate adaptation and stream manipulation require some content to be dropped [6, 12, 3] Furthermore, these schemes tend to be either CPU intensive or require a ....

.... whereas frame dropping techniques take advantage of inter image redundancy [9, 10, 12] An alternate approach is to provide fault resilience by building redundancy into the system as in RAID based servers [2, 10, 9] Rate adaptation and stream manipulation require some content to be dropped [6, 12, 3]. Furthermore, these schemes tend to be either CPU intensive or require a significant time to achieve the desired reduction in resource utilization. RAID based techniques to handle overload typically involve Fault Monitor Aggregation Agent Fault tolerant Storage CM Server Clients Delivery ....

H. M. Vin, P. J. Shenoy and S. Rao, "Efficient Failure Recovery in Multi-Disk Multimedia Servers," Proceedings of the Twenty-Fifth Fault Tolerant Computing Symposium, Pasadena, CA, June 1995, pp. 12-21.

....A single buffer will suffice if a running XOR of the blocks is computed and stored. Multiple parallel read of the parity group could cause deadline violations of continuous media requests. An optimization for relaxing the real time constraints is to pre fetch the blocks 62 in the parity group [7, 120]. When a request is made for a block on the failed disk, the parity group for the next block on the failed disk could be pre fetched and stored in the parity group buffer. In this way no parallel request need be made to read the parity group of the next unavailable block, since they have already ....

....is done in software and is orchestrated by a recovery driver. Details of the techniques used are beyond the scope of this dissertation. 4. 6 Related Work Most previous work on high data availability, examine single node servers and specifically focus on RAID schemes across disks within a node [6, 7, 111, 120]. Several research efforts have focused on guaranteeing availability for continuous media servers. The streaming RAID proposed by Tobagi et al. 111] divides the disks into parity groups 1 . In their scheme, entire parity groups are read out in one cycle and sent on the network during the next ....

H.M. Vin, P.J. Shenoy, and S. S. Rao. Efficient failure recovery in multi-disk multimedia servers. Proceedings of FTCS, June 1995.

....blocks 1 row MPEG stripe Figure 7 Effect of slice allocation on picture quality during failure. 7 Related Work To our knowledge, most previous research has used replication or parity based schemes in designing highlyavailable video servers [2, 8, 11, 13] One exception is the work of Vin et al. [14]. They described a system called Self Recovering Array of Disks (SRAD) to provide highly available video access in the context of RAID storage systems. SRAD modifies the JPEG compression algorithm to split the picture into several subpictures in the frequency domain. The subpictures are stored on ....

H.M. Vin, P.J. Shenoy, and S. Rao. Efficient failure recovery in multi-disk multimedia servers. In Proceedings of the 25th International Symposium on Fault-Tolerant Computing, pages 12--21, June 1995.

.... Fundamentally, both techniques achieve fault resilience by reducing the resources required by a video stream (measured by its bandwidth) Content alteration techniques take advantage of the intra image redundancy whereas frame dropping techniques take advantage of inter image redundancy [8, 9, 10]. An alternate approach is to provide fault resilience by building redundancy into the system as in RAID based servers [2, 9, 8] Rate adaptation and stream manipulation require some content to be dropped [6, 10, 3] Furthermore, these schemes tend to be either CPU intensive or require a ....

.... whereas frame dropping techniques take advantage of inter image redundancy [8, 9, 10] An alternate approach is to provide fault resilience by building redundancy into the system as in RAID based servers [2, 9, 8] Rate adaptation and stream manipulation require some content to be dropped [6, 10, 3]. Furthermore, these schemes tend to be either CPU intensive or require a significant time to achieve the desired reduction in resource utilization. RAID based techniques to handle overload typically involve over allocation of resources. In large systems, introducing redundancies for correcting ....

H. M. Vin, P. J. Shenoy and S. Rao, "Efficient Failure Recovery in Multi-Disk Multimedia Servers," Proceedings of the Twenty-Fifth Fault Tolerant Computing Symposium, Pasadena, CA, June 1995, pp. 12-21.

.... are proposed in [8] on which a clustered architecture can be built [7] On line failure recovery algorithms can allow a fast recovery process by either absorbing the disk bandwidth not consumed by the user processes [5] or by utilizing the inherent redundancy in video streams of the application [14]. We consider fault tolerance in a system of communicating processors that work together to provide some service. An example of such a system might be a cluster of workstations providing web service to a client. The client requests a page and the server machines respond by delivering that page. A ....

H.M. Vin, P.J. Shenoy, and S. Rao, "Efficient Failure Recovery in Multi-Disk Multimedia Servers," in Proc. 25th International Symposium on FaultTolerant Computing (FTCS-25), June 1995, pp. 1221.

No context found.

H. M. Vin, P. J. Shenoy, and S. Rao. Efficient Failure Recovery in Multi-Disk Multimedia Servers. In Proceedings of the 25th International Symposium on Fault Tolerant Computing Systems, Pasadena, CA, pages 12--21, June 1995.

....Continuous media applications with stringent requirements also require such exact reconstruction of lost data. However, for many continuous media applications approximate reconstruction of lost data may be sufficient, so long as the reconstructed data is a reasonable approximation of the original [33]. Such recovery techniques exploit the spatial and temporal redundancies present within continuous media files to approximately reconstruct data stored on the failed disk. Unlike mirroring and parity based techniques, these techniques do not require any additional data to be accessed for failure ....

.... layer supports two mechanisms: 1) a reliable read, in which parity information is used to reconstruct blocks stored on the failed disk, and (2) an unreliable read, in which parity based reconstruction is disabled, thereby shifting the responsibility of failure recovery to the client [33]. Unlike read requests, parity computation can not be disabled while writing blocks. This is because, it may not be known whether reliable reads or unreliable reads (or both) would be used to access the blocks being written. Even if this information is known a priori, parity information must still ....

[Article contains additional citation context not shown here]

H. M. Vin, P. J. Shenoy, and S. Rao. Efficient Failure Recovery in Multi-Disk Multimedia Servers. In Proceedings of the 25th International Symposium on Fault Tolerant Computing Systems, Pasadena, CA, pages 12--21, June 1995.

....the overhead of the on line failure recovery process. Similarly, since human perception is tolerant to minor distortions in media playback, a multimedia server can reduce the overhead of failure recovery by approximating lost data using spatial and temporal redundancies inherent in video streams [4]. That is, a server can partition each image in the video stream into several sub images such that the information contained in each sub image can be approximated using the other sub images. Observe that, such a partitioning process, if carried out in the pixel domain, can substantially reduce ....

H. M. Vin, P. J. Shenoy, and S. Rao. Efficient Failure Recovery in Multi-Disk Multimedia Servers. In Proceedings of the 25th International Symposium on Fault Tolerant Computing Systems, Pasadena, CA, pages 12--21, June 1995.

....the ability to reuse primitive modules (e.g. DCT and huffman) provides the necessary flexibility to create variations of complex modules. In fact, we find this capability to be very useful for constructing prototype codecs that support our research in fault tolerant storage and transmission [8, 14]. 3 PPE Composition: Examples This section describes the construction of PPEs using the library of reusable modules. We first develop a simple JPEG [15] decoder and then demonstrate that, due to a significant functional overlap, an MPEG 2 [9] Spatial SNR scalable decoder can be constructed by ....

H. M. Vin, P. J. Shenoy, and S. Rao. Efficient Failure Recovery in Multi-Disk Multimedia Servers. In Proceedings of the 25th International Symposium on Fault Tolerant Computing Systems, Pasadena, CA, June 1995.

No context found.

H. M. Vin, P. J. Shenoy, and S. Rao. Efficient Failure Recovery in Multi-Disk Multimedia Servers. In Proceedings of the 25th International Symposium on Fault Tolerant Computing Systems, Pasadena, CA, June 1995.

....block from each of the surviving (G Gamma 1) disks in the parity group. Assume that the placement strategy used by the server to interleave media blocks on the array ensures that the parity information is always computed over a sequence of (G Gamma 1) media blocks from the same video stream [33, 34]. That is, all (G Gamma 1) media blocks within a parity group are successive blocks of the same video stream. In such a scenario, due to sequentiality of video playback, a subset of the (G Gamma 1) blocks (namely, blocks from disks (i 1) through G) accessed for recovering the lost block will ....

H M. Vin, P J. Shenoy, and S Rao. Efficient Failure Recovery in Multi-disk Multimedia Servers. In Proceedings of the SPIE Symposium on Voice, Video and Data Communications, Conference on High-Density Data Recording and Retrieval Technologies,Volume 2604, October 1995.

....block from each of the surviving (G Gamma 1) disks in the parity group. Assume that the placement strategy used by the server to interleave media blocks on the array ensures that the parity information is always computed over a sequence of (G Gamma 1) media blocks from the same video stream [33, 34]. That is, all (G Gamma 1) media blocks within a parity group are successive blocks of the same video stream. In such a scenario, due to sequentiality of video playback, a subset of the (G Gamma 1) blocks (namely, blocks from disks (i 1) through G) accessed for recovering the lost block will ....

H. M. Vin, P. J. Shenoy, and S. Rao. Efficient Failure Recovery in Multi-Disk Multimedia Servers. In Proceedings of the 25th International Symposium on Fault Tolerant Computing Systems, Pasadena, CA, pages 12--21, June 1995.

....service times (i.e. the total time spent in retrieving images during a round) that exceed T , resulting in playback discontinuities at client sites. Alternatively, consider a multimedia server that always computes parity over a sequence of (G Gamma 1) media blocks from the same video stream [32]. That is, all (G Gamma 1) media blocks within a parity group are successive blocks of the same video stream. In such a scenario, the server can exploit sequentiality of media playback to minimize the increase in load due to a failure by using media blocks retrieved during playback for failure ....

H. M. Vin, P. J. Shenoy, and S. Rao. Efficient Failure Recovery in Multi-Disk Multimedia Servers. In Proceedings of the 25th International Symposium on Fault Tolerant Computing Systems, Pasadena, CA, pages 12--21, June 1995.

....the following theorem: Theorem 1 To ensure that none of the blocks contained in a sub image are in the 8 neighborhood of each other in the original image, the image must be partitioned into 4 sub images. Proof: The necessity and the sufficiency of this condition can be derived by construction [17]. Notice that when an image is partitioned into 4 subimages, each sub image contains 25 of the image data in the frequency domain. Consequently, if the information contained in a sub image is not available, the image will have to be reconstructed from the remaining 75 of the data. Since the ....

H. M. Vin, P. J. Shenoy, and S. Rao. Efficient Failure Recovery in Multi-Disk Multimedia Servers. Technical Report TR-95-15, Department of Computer Sciences, Univ. of Texas at Austin, 1995.

No context found.

H. M. Vin, P. J. Shenoy,and S. Rao, "Efficient failure recovery in multi-disk multimedia servers," in Proceedingsof the 25th International Symposium on Fault Tolerant Computing Systems, Pasadena, CA, pp. 12--21, June 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