Peter M. Chen, Garth A. Gibson, Randy H. Katz, and David A. Patterson. An evaluation of redundant arrays of disks using an Amdahl 5890. In Proc. SIGMETRICS, pages 74--85, May 1990.  Home/Search   Document Details and Download   Summary   Related Articles   Check

.... (tens to hundreds of terabytes) and high availability (achieved by a combination of redundant storage, support for hot swapping, and transparent fail over capabilities) The internal architectures of commercial disk arrays have become considerably more complex than the early academic prototypes [5] built a decade ago. Array internals now include hardware and firmware support for a variety of optimizations such as adaptive prefetching policies, automatic detection of sequential streams of accesses, efficient demotion policies of dirty blocks from cache, and request coalescing. The presence ....

P.M. Chen, G.A. Gibson, R.H. Katz, and D.A. Patterson. An evaluation of redundant arrays of disks using an Amdahl 5890. In Proc. of ACM Conf. on Measurement and Modeling of Computer Systems (SIGMETRICS), pages 74--85, May 1990.

....number of small inexpensive disks. There exist several alternative ways to configure a number of disks to achieve high performance. We will consider disk striping, data interleaving and combinations of both [RB89a] The RAID project developed at UC Berkeley [KOPS88, CGK 88, KGP89, KOP 89, CGKP90, CK91] incorporates an array of small inexpensive disks. The problem of reliability is solved by using extra disks to store redundant information. This approach CHAPTER 2. DESIGN ISSUES IN PARALLEL I O SYSTEMS 8 is called low level parallelism by Ghosh et al. Joy93] The RAID system breaks the ....

Peter Chen, Garth Gibson, Randy Katz, and David Patterson. An Evaluation of Redundant Arrays of Disks using an Amdahl 5890. In Proceedings of the 1990 ACM Sigmetrics Conference on Measurement and Modeling of Computer Systems, pages 74--85, May 1990.

....taxonomy. In RAID Level 5 picture, blocks labeled Pn are parity blocks. 5.2.3 RAID Taxonomy The RAID group at U. C. Berkeley developed a taxonomy for RAID systems to describe the different options for data interleaving and redundancy [19] Figure 5. 3 shows several RAID classifications or levels [18]. RAID Level 0 is a non redundant collection of disks. RAID Level 1 uses mirroring; whenever data are written to a disk, a copy of the data is written to a redundant disk. Mirrored systems contain twice as many disks as non redundant disk arrays. RAID Level 2 decreases the number of disks required ....

Peter M. Chen, Garth A. Gibson, Randy H. Katz, and David A. Patterson. An evaluation of redundant arrays of disks using an amdahl 5890. In Proceedings SIGMETRICS, pages 74--85, May 1990.

....space management algorithms. Detailed cost performance tradeoffs of such an approach is beyond the scope of this paper. 4 Related Work Several research projects have developed simulation and analytical techniques for optimizing the performance of striped disk arrays for conventional workloads [3, 4, 5, 14]. As demonstrated in Section 1, due to the real time nature of continuous media accesses, these techniques are not directly applicable for optimizing performance in multimedia servers. The problem of determining the optimal stripe unit size for non redundant arrays storing continuous media was ....

P. M. Chen, G. A. Gibson, R. H. Katz, and D. A. Patterson. An Evaluation of Redundant Arrays of Disks using an Amdahl 5890. In Proceedings of ACM SIGMETRICS, pages 74--85, May 1990.

....space management algorithms. Detailed cost performance tradeoffs of such an approach is beyond the scope of this paper. 4 Related Work Several research projects have developed simulation and analytical techniques for optimizing performance of striped disk arrays in conventional file servers [3, 4, 5, 14]. However, as illustrated in Section 1, due to the real time nature 1360 1380 1400 1420 1440 1460 1480 0 20 40 60 80 100 120 Partition size (a) Number of clients supported vs partition size, disks=120 Number of clients 0 2 4 6 8 10 12 14 0 20 40 60 80 100 120 Array Size (b) Optimal parition ....

P. M. Chen, G. A. Gibson, R. H. Katz, and D. A. Patterson. An Evaluation of Redundant Arrays of Disks using an Amdahl 5890. In Proceedings of ACM SIGMETRICS, pages 74--85, May 1990.

....space management algorithms. Detailed cost performance tradeoffs of such an approach is beyond the scope of this paper. 4 Related Work Several research projects have developed simulation and analytical techniques for optimizing the performance of striped disk arrays for conventional workloads [3, 4, 5, 14]. As demonstrated in Section 1, due to the real time nature of continuous media accesses, these techniques are not directly applicable for optimizing performance in multimedia servers. The problem of determining the optimal stripe unit size for non redundant arrays storing continuous media was ....

P. M. Chen, G. A. Gibson, R. H. Katz, and D. A. Patterson. An Evaluation of Redundant Arrays of Disks using an Amdahl 5890. In Proceedings of ACM SIGMETRICS, pages 74--85, May 1990.

....large amounts of memory. However, it is not clear how to scale I O capabilities to keep pace with FLOPS; there are no scalable I O systems that eliminate the I O bottleneck for parallel computers. Disk Array (RAID) technology is one of the enabling hardware technologies for parallel I O [11, 10, 30]. However, in most cases, RAID provides the traditional interface of a single logical disk where the parallelism is hidden at the lower (hardware level) Furthermore, disk arrays only address small scale parallelism and do not directly address the important software issues including parallel file ....

Peter Chen, Garth Gibson, Randy Katz, and David Patterson. An evaluation of redundant arrays of disks using an Amdahl 5890. In Proceedings of the 1990 ACM Sigmetrics Conference on Measurement and Modeling of Computer Systems, pages 74--85, May 1990.

....if a large number of disks are concurrently transferring data over the network [56] 4 Hence, it is important that most of a processor s I O requests be directed to nearby devices. An alternative to distributing the disks across the multiprocessor is to put all disks in a single large disk array [7, 19, 20, 23, 47, 62, 101, 102, 107], where the distribution of data to the disks is handled by a hardware controller associated 3 In the context of HFS, when we refer to the file system software we are referring to all components of the file system including the servers, application level library, and device drivers. 4 This is ....

Peter Chen, Garth Gibson, Randy Katz, and David Patterson. An evaluation of redundant arrays of disks using an Amdahl 5890. In Proceedings of the 1990 ACM Sigmetrics Conference on Measurement and Modeling of Computer Systems, pages 74--85, May 1990.

....provide. Because data is placed under two independent disks, we can further improve read rate by a technique called seek scheduling, which schedules the disk with its head closer to the requested data to service the request. Fifteen to twenty percent improvement to read rate has been observed [10]. Assuming a reasonably short repair time, such as 72 hours, the probability that both members of a mirrored pair fail at the same time is very small, and thus a very large expected mean time to data loss (MTTDL) can be achieved. 111 0 n 1 n 1 n 1 P P 2n 1 P Figure 5.1: RAID Level 5: Parity ....

P. Chen, G. Gibson, R. Katz, and D. Patterson. An evaluation of redundant arrays of disks using amdahl 5890. In ACM SIGMETRICS Conference, Boulder, CO, May 1990.

....the system s behavior during normal operation and under failure as well as its ability to recover quickly and return to the fully operational state. When designing a fault tolerance scheme, the following aspects of the disk subsystem must be examined: a) performance under normal operation (e.g. [7, 31]) b) mean time to data loss (or system failure) e.g. 14] and c) performance of the disk subsystem under failure, i.e. when one or more disks are inoperable or inaccessible (e.g. 35, 30, 17, 19, 18] We should keep in mind that failures are expected to occur relatively infrequently, so ....

P. Chen, G. A. Gibson, R. H. Katz, and D. A. Patterson. An evaluation of redundant arrays of disks using an Amdahl 5890. ACM SIGMETRICS Conference, pages 74--85, 1990.

....to rebuild a failed disk (and therefore the MTTDL) and, b) the workload (measured in accesses per second per disk) that can be supported during the rebuild process, and c) the system s performance under failure. The MTTDL of a RAID is easily shown to be inversely proportional to the rebuild time [6, 11]; in the RAID system described in [31] rebuilding the failed disk contents at maximum speed (the capacity of the standby disk) results in the use of the entire capacity of the surviving disks in the array. Thus rebuilding at maximum rate means that the array can perform no other work during the ....

P. Chen. An evaluation of redundant arrays of disks using an Amdahl 5890. Technical Report UCB/CSD 89/506, UC Berkeley, May 1989.

No context found.

Peter M. Chen, Garth A. Gibson, Randy H. Katz, and David A. Patterson. An evaluation of redundant arrays of disks using an Amdahl 5890. In Proc. SIGMETRICS, pages 74--85, May 1990.

....disks as a nonredundant disk array [Bitton88] Whenever data is written to a disk the same data is also written to a redundant disk, so that there are always two copies of the information. When data is read, it can be retrieved from the disk with the shorter queueing, seek and rotational delays [Chen90a]. If a disk fails, the second copy is used to service requests. Mirroring is frequently used in database applications where availability and transaction rate are more important than storage efficiency [Gray90] 1. Strictly speaking, RAID Level 0 is not a type of redundant array of inexpensive ....

Peter M. Chen, Garth Gibson, Randy H. Katz, and David A. Patterson. An Evaluation of Redundant Arrays of Disks Using an Amdahl 5890. In Proceedings of the 1990 ACM SIGMETRICS Conference on Measurement and Modeling of Computer Systems, May 1990. The first experimental evaluation of RAID. Compares RAID levels 0, 1, and 5.

....block is not aligned with the stripe boundary, a partial stripe write occurs and a read update procedure has to be followed. For simplicity, we assume that a partial stripe write is only required for the last stripe of each I O, i.e. the starting block is always aligned with the stripe boundary [5]. This may yield an optimistic estimate of the performance of RAID 5. Let p f be the probability that a write request is a full stripe write. For a partial stripe write, since it accesses multiple blocks from each disk, after reading out a block from the last stripe (for the purpose of a new ....

Chen, P., Gibson, G., Katz, R. H., and Patterson, D. A. An evaluation of redundant arrays of disks using an Amdahl 5890. Perf. Eval. Rev., 18(1):74--85, May 1990.

No context found.

Peter M. Chen, Garth A. Gibson, Randy H. Katz, and David A. Patterson. An evaluation of redundant arrays of disks using an Amdahl 5890. Proceedings of SIGMETRICS (Boulder, Colorado), pages 74--85, 22--25 May 1990.

No context found.

P. Chen. An evaluation of redundant arrays of disks using an Amdahl 5890. Technical Report UCB/CSD 89/506, Computer Science Division, University of California, Berkeley, May 1989.

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