T. Kimbrel and A. R. Karlin. Nearoptimal parallel prefetching and caching. In Proc. of the IEEE Symposium on Foundations of Computer Science, 1996.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....optimal performance. Rajasekaran [22] gives another asymptotically suboptimal deterministic parallel disk sorting algorithm that runs in three passes for not too large inputs. Prefetch bu#ers for disk load balancing and overlapping of I O and computation has been intensively studied [21, 7, 3, 14, 13, 12]. But we have not seen results that guarantee overlapping of I O and computation during parallel disks merging of arbitrary runs. There are many good practical implementations of sorting (e.g. 19, 1, 30, 20] that address parallel disks, overlapping of I O and computation, and low internal ....

T. Kimbrel and A. R. Karlin. Near-optimal parallel prefetching and caching. SIAM Journal on Computing, 29(4):1051--1082, 2000.

....of reference patterns, Figure 3.1 demonstrated that the leakage energy to be saved would be significant. The question is: can we develop policies that come acceptably close to this oracle In fact, this question can be approached by relating it to the theoretical area of competitive algorithms [44]. Competitive algorithms make cost benefit decisions online (i.e. without oracle knowledge of the future) that offer benefits within a constant factor of an optimal offline (i.e. oracle based) algorithm. A body of computer systems work has previously successfully applied such strategies to ....

T. Kimbrel and A. Karlin. Near-Optimal Parallel Prefetching and Caching. SIAM Journal on computing, 2000.

....model. Cao et al. 7] propose a model that additionally allows overlapping of I O and computation. Albers et al. 2] were the first to find an optimal polynomial time offline algorithm for the single disk case in this model but it does not generalize well to multiple disks. Kimbrel and Karlin [12] devised a simple algorithm called reverse aggressive that obtains good approximations in the parallel disk case if the buffer pool is large and the failure penalty F is small. However, in our model, which corresponds to F 1, the approximation ratio that they show goes to infinity. Reverse ....

Tracy Kimbrel and Anna R. Karlin. Near-optimal parallel prefetching and caching. SIAM Journal on Computing, 29(4):1051--1082, 2000.

....data block, it should be resident in the internal memory of the computer system. By serving a reference string, we refer to the act of carrying out a series of I O operations that make it possible for the computation to access blocks in the order specified by the reference string. A recent study [6] focussed on the off line problem of serving an arbitrary but fully known reference string of blocks spread across D parallel, independent disks using parallel prefetching in conjunction with page replacement . The authors presented and analyzed a very clever but somewhat complicated ....

....more tightly coupled system where the buffer can be shared by the different disks was not considered. In this paper we present a competitive analysis framework for parallel prefetching algorithms on parallel disk systems for a restricted family of reference strings. In contrast to the requirement [6] of knowing a priori the entire reference string exactly, our parallel prefetching approach is based on models of bounded lookahead that are easily realizable in practice. Our restricted family of reference strings are called read once consumption sequences, in which all references are read only ....

T. Kimbrel and A. R. Karlin. Near-Optimal Parallel Prefetching and Caching. In 37th Annual Symposium on Foundations of Computer Science, October 1996.

....model. Cao et al. 7] propose a model that additionally allows overlapping of I O and computation. Albers et al. 2] were the first to find an optimal polynomial time offline algorithm for the single disk case in this model but it does not generalize well to multiple disks. Kimbrel and Karlin [12] devised a simple algorithm called reverse aggressive that obtains good approximations in the parallel disk case if the buffer pool is large and the failure penalty F is small. However, in our model, which corresponds to F 1, the approximation ratio that they show goes to infinity. Reverse ....

Tracy Kimbrel and Anna R. Karlin. Near-optimal parallel prefetching and caching. SIAM Journal on Computing, 29(4):1051--1082, 2000.

....which every optimal single disk algorithm must conform. One of these principles is what we call FIFO in this paper. Albers et al. 2] were the first to find an optimal polynomial time offline algorithm for the single disk case but it does not generalize well to multiple disks. Kimbrel and Karlin [12] devised a simple algorithm called reverse aggressive that obtains good approximations in the parallel disk case if the buffer pool is large and the failure penalty F is small. However, in our model, which corresponds to F 1 the approximation ratio that 3 they show goes to infinity. Kallahalla ....

....for the reverse sequence Sigma R . This immediately implies an optimal, linear time algorithm for prefetching: greedyWriting mutates into lazyPrefetching via the duality, which implies its optimality as a prefetching algorithm. Lazy prefetching is equivalent to the reverse aggressive algorithm [12] for the case of read once sequences and proves that the latter is optimal in the case where the failure penalty F goes to infinity. A practical disadvantage of lazyPrefetching is its laziness; it delays fetching a block as long as possible. Barve et al. 4] and Kallahalla and Varman [10] give ....

Tracy Kimbrel and Anna R. Karlin. Near-optimal parallel prefetching and caching. SIAM Journal on Computing, 29(4):1051--1082, 2000.

....an optimal, linear time algorithm for prefetching: for read sequences consisting of distinct blocks greedyWriting mutates into lazyPrefetching via the duality, which also implies its optimality as a prefetching algorithm. Algorithm lazyPrefetching is equivalent to the reverse aggressive algorithm [5] for the case of read once sequences and proves that the latter is optimal in the case where the failure penalty F goes to infinity. 1 A practical disadvantage of lazyPrefetching is its laziness; it delays fetching a block as long as possible. Barve et al. 1] and Kallahalla and Varman [3] give ....

Kimbrel, T., and Karlin, A. R. Near-optimal parallel prefetching and caching. SIAM Journal on Computing 29, 4 (2000), 1051--1082.

....Cao et al. 5] this paper studies offline prefetching and caching algorithms, where fetches are serialized, i.e. at most one fetch can be in progress at any given time. An algorithm, called Aggressive (AGG) was shown to yield a min(1 d k ; 2) approximation to the optimal. Karlin and Kimbrel [13] extended this study to storage systems which consist of r units (e.g. disks) fetches are serialized on each storage unit, thus, up to r block fetches can be processed in parallel. The paper gives performance bounds for several offline algorithms in this setting. Algorithm AGG is shown to ....

A. R. Karlin, T. Kimbrel, "Near-optimal parallel prefetching and caching", CS TR, Washington Univ., 1996.

....of reference patterns, Figure 2 demonstrated that the leakage energy to be saved would be significant. The question is: can we develop policies that come acceptably close to this oracle In fact, this question can be approached by relating it to the theoretical area of competitive algorithms [19]. Competitive algorithms make cost benefit decisions online (i.e. without oracle knowledge of the future) that offer benefits within a constant factor of an optimal offline (i.e. oracle based) algorithm. A body of computer systems work has previously successfully applied such strategies to ....

T. Kimbrel and A. Karlin. Near-optimal parallel prefetching and caching. SIAM Journal on computing, 2000.

....read many reference strings where the same block may be referenced more than once. Other formal studies have looked at the joint problem of prefetching and caching general reference strings (with repetitions) in other parallel I O models. A parallel version of the stall model was introduced in [10]. A sophisticated off line approximation algorithm was analyzed and shown to be efficient with respect to elapsed time, for typical system parameters (memory size and ratio of I O to computation time) A distributed buffer system, a variant of the parallel disk model, where each disk has its ....

T. Kimbrel and A. R. Karlin. Near-optimal parallel prefetching and caching. In Proc. of 37th Annual Symp. on Foundations of Computer Science, pages 540--549. IEEE, October 1996.

....string since there is no reuse of blocks. For general read many reference strings, in the stall model of parallel I O (a generalization of the sequential model of [6] to multiple disks and a shared buffer) an approximate off line algorithm for I O scheduling and buffer management was presented in [10]. An optimal offline buffer management and scheduling algorithm was presented in [14] for a distributedbuffer parallel I O model in which each disk has its own private buffer. 2 Definitions We use the Parallel Disk Model [16] consisting of D independent disks and an M block I O buffer, M D. ....

T. Kimbrel and A. R. Karlin. Near-optimal parallel prefetching and caching. In Proc. of 37th Annual Symp. on Foundations of Computer Science, pages 540--549. IEEE, October 1996.

....Second, many systems employ new techniques to proactively manage movement and placement of data in the storage hierarchy. In particular, prefetching is now a part of every advanced I O system, and recent research has yielded a family of integrated prefetching and caching schemes [Cao et al. 1995; Kimbrel et al. 1996; Patterson et al. 1995] The most flexible and powerful tool for evaluating new I O structures is simulation, which handles varying assumptions about the workload and the hardware. Unfortunately, high overhead compromises the usefulness of simulation. While execution driven simulation eliminates ....

....reduction for I O caching systems that proactively manage a storage hierarchy, particularly for systems that use prefetching. We are not aware of any previous trace reduction work that addresses the prefetching aspect, despite the volume of research in prefetching algorithms [Cao et al. 1995; Kimbrel et al. 1996; Patterson et al. 1995] and prefetching for virtual memory [Trivedi 1976; Mowry et al. 1996; Voelker et al. 1998] This paper presents a prefetch safe trace reduction algorithm, called FastSlim, that yields exact simulations for a large class of prefetching schemes with integrated caching and ....

[Article contains additional citation context not shown here]

Kimbrel, T. and Karlin, A. R. 1996. Near-optimal parallel prefetching and caching. In 37th Annual Symposium on Foundations of Computer Science (Burlington, Vermont, 14--16 Oct. 1996), pp. 540--549.

....(and demand paging) algorithms. However, with respect to the class of all prefetching algorithms, optimality in terms of the number of page faults does not guarantee the optimality in terms of the total I O stall time. More recently, new approaches to caching, paging and prefetching (see [10, 11] for example) based on the framework of competitive analysis [12] have led to new insights and practical improvements. In order to facilitate specifically the design and analysis of prefetching algorithms, Cao et al. 10] proposed the elegant stall time model, in which accessing a page in cache ....

....F K g. They also presented a simple argument proving that R MIN and R Conservative are at most 2: This argument 2 hinges on the fact that the total number of page fetch operations issued by Conservative and MIN on any request sequence is the smallest possible. Subsequently, Kimbrel and Karlin [11] proposed the Reverse Aggressive prefetching algorithm, whose total elapsed time is within a factor of 1 d(F 1) K of the optimal elapsed time, when there are d disks that may be used in parallel. Very recently, Albers, Garg and Leonardi [13] used sophisticated techniques to devise an optimal ....

[Article contains additional citation context not shown here]

T. Kimbrel and A. R. Karlin, "Near-optimal parallel prefetching and caching," in 37th Annual Symposium on Foundations of Computer Science, (Burlington, Vermont), pp. 540--549, 14--16 Oct. 1996.

....is needed in order to prefetch effectively as well. The study of lookahead models for parallel I O systems was introduced in [3] for read once reference strings. An approximate offline deterministic prefetching and buffer management algorithm for the stall model of parallel I O was presented in [11]. An optimal offline deterministic algorithm for the distributed buffer parallel I O system was presented in [15] 1.1 Definitions The notions of read once and read many reference strings and lookahead, which we introduced informally in the previous section, are defined below. Definition 1. The ....

Kimbrel, T., and Karlin, A.: Near Optimal Parallel Prefetching and Caching. 37th Ann. Symp. on Foundations of Computer Science (1996)

....parallel I O model in which each disk has its own private buffer. In an alternative stall model of parallel I O, a generalization of the sequential model of [6] to multiple disks and a shared buffer, an approximate off line algorithm for I O scheduling and buffer management was presented in [10]. However, so far the question of devising an on line algorithm with bounded lookahead for general read many reference strings in a parallel I O model has not been addressed. In this paper we study the on line I O scheduling problem for read many reference strings in the framework of competitive ....

T. Kimbrel and A. R. Karlin. Near-Optimal Parallel Prefetching and Caching. In Proc. of FOCS'96, pages 540--549. IEEE, Oct. 1996.

....which stripes data starting from a randomly chosen disk. These results are closest to our result on scheduling read once reference strings using a randomized data layout. In contrast the work here deals with generalized reference strings. In a generalization of the stall model to parallel disks, [28, 27] focussed on the off line problem of serving read many reference strings. They designed a sophisticated off line approximation algorithm called reverse aggressive to service read many reference strings. Reverse aggressive is shown to be near optimal for typical system parameters (memory size and ....

T. Kimbrel and A. R. Karlin. Near-optimal parallel prefetching and caching. In Proc. of 37th Annual Symp. on Foundations of Computer Science, pages 540--549. IEEE, October 1996.

....to overlap CPU and I O operations was addressed in [6] using a stall model of I O. Off line approximation algorithms to minimize the stall time were presented and analyzed. In the multiple disk situation, previous studies have either designed off line algorithms in other parallel I O models [12, 20], or made assumptions about the structure of the access patterns [16, 3, 11] or designed parallel prefetching algorithms in the context of specific applications [17, 18, 14, 15, 2] However, so far the question of devising an on line algorithm with bounded lookahead for arbitrary read many ....

....context of specific applications [17, 18, 14, 15, 2] However, so far the question of devising an on line algorithm with bounded lookahead for arbitrary read many reference strings in a parallel I O model has not been addressed. The stall model of [6] was generalized for multiple disk systems in [12]. A sophisticated off line approximation algorithm called reverse aggressive for parallel disk scheduling was presented and analyzed. Its competitive ratio was shown to be small for typical values of CPU and disk speeds, buffer size, and number of disks. Disk scheduling in a distributedbuffer ....

T. Kimbrel and A. R. Karlin. Near-Optimal Parallel Prefetching and Caching. In 37th Annual Symposium on Foundations of Computer Science, pages 540--549. IEEE, October 1996.

....system s behavior such as caching strategy and file layout, and takes into account the behavioral characteristics of the disks used to store files. Third, our model predicts the performance of the file system. Substantial work has been done studying the interaction between prefetching and caching [Cao95, Patterson95, Kimbrel96]. Others have examined methods to work around the file system cache to achieve the desired performance (e.g. Kotz95] The benefit of prefetching is not limited to workloads where files are read sequentially; Small studied the effect of prefetching on random access, zero think time workloads on ....

Tracy Kimbrel and Anna R. Karlin. Near-optimal parallel prefetching and caching. Proceedings of the 37th Annual Symposium on Foundations of Computer Science (Burlington, VT), pages 540--549, October 1996.

....focuses primarily on prefetching algorithms which work with a statically scheduled synchronization algorithm. The static schedule produced gives the prefetching algorithm perfect knowledge of future memory references. A number of papers analyze prefetching algorithms which assume such knowledge [3, 13, 37, 38, 72]. Tomkins summarizes the results in his thesis [72] In the model used in these papers one unit of time is spent doing computation between each reference. A disk fetch, including a write of the block being replaced and a read of the block fetched, takes F units of time. Memory holds m blocks. In ....

....algorithm performs better than the conservative algorithm. We believe their proof of the bounds is incorrect. We have an example of a schedule, not shown here, for which a statement made in their proof is false. We are currently working on a corrected proof of the same result. Kimbrel and Karlin [37] suggest another algorithm, reverse aggressive, which has provably good better performance than aggressive on multi disk systems, where each block is on one of several disks. This algorithm has also shown empirically better performance than aggressive on single disk systems. In the single disk ....

Tracy Kimbrel and Anna R. Karlin. Near-optimal parallel prefetching and caching. In Proceedings of the 1996 IEEE Symposium on Foundations of Computer Science, October 1996.

....block to replace whose next reference is after the request to the missing block. 3. The Controlled Aggressive Strategy. The controlled aggressive strategy behaves like the aggressive strategy but also considers the disk workload. A prefetch is issued only when the disk is idle. Kimbrel and Karlin [Kimbrel et al. 1996a; Kimbrel and Karlin, 1996; Kimbrel et al. 1996b; Kimbrel, 1997] extended the work of Cao et al. for parallel disk access with two prefetch algorithms: 1. The Reverse Aggressive Algorithm. At first the algorithm transforms the reference string into the reverse reference string. Switching between ....

....reference is after the request to the missing block. 3. The Controlled Aggressive Strategy. The controlled aggressive strategy behaves like the aggressive strategy but also considers the disk workload. A prefetch is issued only when the disk is idle. Kimbrel and Karlin [Kimbrel et al. 1996a; Kimbrel and Karlin, 1996; Kimbrel et al. 1996b; Kimbrel, 1997] extended the work of Cao et al. for parallel disk access with two prefetch algorithms: 1. The Reverse Aggressive Algorithm. At first the algorithm transforms the reference string into the reverse reference string. Switching between forward and reverse ....

[Article contains additional citation context not shown here]

Kimbrel, T. and Karlin, A. (1996). Near-optimal Parallel Prefetching and Caching.

....latency. The tension between caching and prefetching and its associated tradeoffs are particularly important in Internet application environments, as the potential benefits of such methods can be quite significant. We note that optimal algorithms have been developed for these types of problems [Kimbrel and Karlin 1996]. These algorithms, however, require the availability of some form of advance knowledge of future data requests. Flintstone can be used to provide accurate estimates of this information for each Web page by exploiting aspects of our methodology in ways that have been proven effective in other ....

Kimbrel, T. and A. R. Karlin (1996), "Near-optimal parallel prefetching and caching," In Proceedings of the IEEE Symposium on Foundations of Computer Science.

....new model for integrated prefetching and caching was introduced. In this model the time to consume a block is an explicit parameter, and the elapsed time is the performance metric. Off line approximation algorithms for a single disk and multiple disk systems in this model were addressed in [5] and [9] respectively. Recently a polynomial time optimal algorithm for the single disk case was discovered in [1] In the parallel disk model used in this paper, a randomized caching and scheduling algorithm with bounded expected performance was presented in [8] Using a distributed buffer configuration, ....

T. Kimbrel and A. R. Karlin. Near-Optimal Parallel Prefetching and Caching. In Proc. of FOCS'96, pages 540-- 549. IEEE, October 1996.

....a multiple disk environment were studied in [10] using Markov analysis for small configurations and simulation for larger systems. Both of these works assumed uniformly random data (no skew) in the evaluation. Off line prefetching algorithms in the global and distributed models were presented in [7] and [16] respectively. Competitive analysis of on line prefetching with bounded lookahead has been analyzed in [3, 6] for the global buffer model, and [16] for the distributed buffer model. Other works on improving the performance of external sorting and merging using a single disk system include ....

T. Kimbrel and A. R. Karlin. Near-Optimal Parallel Prefetching and Caching. In 37th Annual Symposium on Foundations of Computer Science, pages 540--549. IEEE, October 1996.

No context found.

T. Kimbrel and A. R. Karlin. Nearoptimal parallel prefetching and caching. In Proc. of the IEEE Symposium on Foundations of Computer Science, 1996.

No context found.

Kimbrel, T., and Karlin, A.: Near Optimal Parallel Prefetching and Caching. 37th Ann. Symp. on Foundations of Computer Science #1996#

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