D.L. Willick, D.L. Eager, and R.B. Bunt, "Disk Cache Replacement Policies for Network Fileservers," Proc. 13th Int'l Conf. Distributed Computing Systems, May 1993.  Home/Search   Document Details and Download   Summary   Related Articles   Check

.... issue of Web cache filter effects, wherein a higher level cache in a multi cache system only handles requests that miss in the lower level cache(s) Similar cache design problems have been addressed previously in the context of CPU cache hierarchies [14] databases [10] and client server systems [16]. Among the papers that focus on Web cache filter effects, most focus on the frequency domain aspect of the Web cache filter effect. Doyle et al. 9] refer to this as the trickle down effect, and conduct a detailed simulation study to quantify its impact. Che et al. 7] propose a ....

D. Willick, D. Eager, and R. Bunt, "Disk Cache Replacement Policies for Network Fileservers", Proceedings of ICDCS, Pittsburgh, PA, 1993.

....locality in Web server traces [2] This work confirms those findings. Furthermore, a variety of advanced replacement policies to exploit this locality have been proposed for the web file cache environment [4, 6] Other work has considered replacement policies for network file server block caches [17]. The specific policy choices for allocation and replacement are independent of the concepts expressed here, so any policies could be adopted. 8. Conclusions Repeatedly transferring frequently requested data across an expensive local interconnect leads to an inefficient use of system resources. ....

Darryl L. Willick, Derek L. Eager, and Richard B. Bunt. Disk Cache Replacement Policies for Network Fileservers. In Proceedings of the 13th International Conference on Distributed Computing Systems, pages 2--11, May 1993.

....by this observation [21, 24, 30, 48] Our DEMOTE scheme, with alternative array cache replacement policies, is another such remedy. Choosing the correct cache replacement policy in an array can improve its performance [19, 21, 30, 35, 48] Some studies suggest using least frequently used [15, 46] or frequency based [30] replacement policies instead of LRU in file servers. MRU [23] or next block prediction [29] policies have been shown to provide better performance for sequential loads. LRU or clocking policies [10] can yield acceptable results for database loads; for example, the IBM DB2 ....

D. L. Willick, D. L. Eager, and R. B. Bunt. Disk cache replacement policies for network fileservers. In Proc. of the 13th Intl. Conf. on Distributed Computing Systems, pages 2--11. May 1993.

....available in [9] In general, the child proxy caches have much higher hit ratios than the parent proxy. This observation is not surprising, given that the parent proxy only sees the requests that miss at the lower level caches (i. e, the request stream is filtered by the lower level proxies) 30] [32]. Figure 2(a) shows that the GD Size policy at the parent cache provides a significantly better document hit ratio than either LRU or LFU Aging at the parent cache. This document hit ratio advantage is a factor of two or more for most of the cache sizes considered. Furthermore, this advantage ....

D. Willick, D. Eager, and R. Bunt, "Disk Cache Replacement Policies for Network Fileservers", Proceedings of ICDCS, Pittsburgh, PA, 1993.

....of a CM object or entire objects. Finally, for environments with highly skewed access patterns (i.e. a large fraction of the accesses go to a disproportionately small set of hot objects [7] instead of caching only portions of a CM object, a server may employ a frequency caching algorithm [20, 24], which caches entire CM objects in cache, and serves subsequent requests for those objects from the cache. Such an algorithm would cache a CM object purely based on its access frequency (and hence, without any considerations of its size) Observe that all of the above caching techniques assume ....

D.L. Willick, D. L. Eager, and R. B. Bunt. Disk cache replacement policies for network fileservers. In Proceedings of International Conference on Distributed Computing Systems, May 1993. 18

....considerably by traditional caching studies (e.g. memory hierarchy caching, distributed le systems) Although the concepts are similar, there are three primary di erences between Web caching and 13 traditional caching. First, traditional caching is often concerned with xed size transfer units [29, 87], whereas Web caching involves variable sized documents, due to the restrictions imposed by the HTTP protocol. Second, documents of the same size often have di erent download times (i.e. di erent associated costs) Third, the Web contains many more les and users than large scale distributed le ....

....Previous Day Days from the Starting Day M = 0.01 M = 0.05 M = 0.10 (b) Figure 3.11: Hot Set Drift for the NLANR Data Set: a) Absolute Drift (b) Relative Drift 54 3.2.8. 2 Short Term Measure of Temporal Locality The Least Recently Used Stack Model (LRUSM) 77] has been widely used [2, 5, 86, 87] to measure temporal locality. The LRUSM is a stack based ordering of referenced objects, according to their recency of reference (i.e. most recently referenced item on top (position 1) and the least recently referenced item on the bottom) For each reference in the request stream, the stack is ....

D. Willick, D. Eager, and R. Bunt, \Disk Cache Replacement Policies for Network File Servers", Proceedings of the 13th International Conference on Distributed Computing Systems (ICDCS), Pittsburgh, PA, pp. 2-11, May 1993.

....rates, because smaller objects tend to have a slightly higher hit rate than larger objects. Figure 1 shows that the hit rate of the NLANR thirdlevel cache is considerably lower than that of the secondlevel caches. This lower hit rate is expected and is a common feature of higher level disk caches [23]. The lowerlevels of a cache hierarchy act as a filter, removing temporal locality and lower level sharing from the reference stream. The resulting references received by the higherlevel cache consist only of the lower level caches capacity misses, consistency misses, and first time requests ....

D. L. Willick, D. L. Eager, and R. B. Bunt. Disk cache replacement policies for network fileservers. In Proceedings of the IEEE International Conference on Distributed Computer Systems, pages 2--11, June 1993.

....the high cost of accessing slow mass storage, even small improvements in hit rate are important. We plan to develop and evaluate caching algorithms tailored to the mass storage and low speed network applications as well as studying caching algorithms that have been suggested by others. Willick [21] and Korner [22] investigate alternatives to LRU caching for a distributed file system environment. We may perform simulations using their suggested caching schemes in the context of multilevel caching. However, it is difficult to guess the effectiveness of these caching policies for our Formatted ....

D. Willick, D. Eager, and R. Bunt, "Disk Cache Replacement Policies for Network Fileservers," pp. 2-11 in Proc. 13th Intl. Conference on Distributed Computing Systems, Pittsburgh (May, 1993).

....rates, because smaller objects tend to have a slightly higher hit rate than larger objects. Figure 1 shows that the hit rate of the NLANR thirdlevel cache is considerably lower than that of the secondlevel caches. This lower hit rate is expected and is a common feature of higher level disk caches [23]. The lowerlevels of a cache hierarchy act as a filter, removing temporal locality and lower level sharing from the reference stream. The resulting references received by the higherlevel cache consist only of the lower level caches capacity misses, consistency misses, and first time requests ....

D. L. Willick, D. L. Eager, and R. B. Bunt. Disk cache replacement policies for network fileservers. In Proceedings of the IEEE International Conference on Distributed Computer Systems, pages 2--11, June 1993.

....cache. The performance improvements resulting from server cache hits, when averaged over a large trace, would effectively result in a slightly lower average remote disk latency for small caches. When client caches are even moderately large (1. 6MB) server cache hit rates drops significantly [WEB93] In addition, our results do not include congestion contention at the server and network that occurs in the real world, which helps offset the missing server cache. Under our remote file system model, it is possible for the prefetching component of the file system to request a file and initiate ....

D. Willick, D. Eager, and R. Bunt. Disk Cache Replacement Policies for Network Fileservers. In Proceedings of the 13th International Conference on Distributed Computing Systems, pages 2--11, May 1993.

....future. Most virtual memory systems actually use an approximation of LRU, as we describe in Section 6.3.4. Less common are frequency based schemes such 17 as LFU (Least Frequently Used) which are less popular because access frequency is more difficult to collect and interpret than access time [55, 82, 34]. There are three well known problems with LRU (and LFU) which are all related to the fact that LRU considers only system wide page access time ( frequency) information and thus does not consider characteristics of the programs that are accessing those pages or the files to which they belong ....

D. L. Willick, D. L. Eager, and R. B. Bunt. Disk cache replacement policies for network fileservers. In Proceedings of the 13th International Conference on Distributed Computing Systems, pages 2--11, may 1993.

....log entries back to the file server when appropriate. This is called trickle reintegration in Coda [2] and background replay in AFS [10] 3. ISSUES IN CLIENT CACHE MANAGEMENT File caching in strongly connected environments is well understood, having been used in distributed systems for many years [16, 17, 18], but the use of file caching to permit disconnected or weakly connected operation is a much more recent innovation. While many file caching issues have been addressed, current systems still have limitations. Ultimately, good management strategies for mobile computing will be those that provide ....

D. Willick, D. Eager, and R. Bunt. Disk cache replacement policies for network file servers. In Proc. Thirteenth International Conference on Distributed Computing Systems, pages 2--11, Pittsburgh, PA, May 1993.

....In a client server environment disk caches can be placed both at the client, to avoid network access to the file server, and at the file server, to avoid physical disk accesses. Such a two level cache environment has been the subject of a number of recent investigations [Makaroff and Eager 90, Willick et al. 93] These studies identified some important differences between the two level cache environment and the more familiar single level cache environment, differences that speak to the suitability of various approaches to cache management. These studies focused only on read performance, however, as ....

....caching policy factor on each of the client and the server were considered: LRU, LRUPT, and WBT. The size of the volatile cache in the single level caching system and on the client and the server of the two level caching system was constant (200 blocks a figure chosen based on the results in [Willick et al. 93] The size of the NV cache is also a factor in the experiment. For the single level caching system, seven levels were considered for the NV cache size (10, 30, 50, 70, 100, 150, and 200 blocks) In the two level caching system, four levels were considered for the NV cache size at the client (0, ....

[Article contains additional citation context not shown here]

Willick, D.L., Eager, D.L., and Bunt, R.B., "Disk Cache Replacement Policies for Network Fileservers", Proceedings of the 13th International Conference on Distributed Computer Systems, Pittsburgh, PA, May 1993, pp.2-11. 22

.... [10] Similar characteristics have been observed for le reference patterns [18, 21] Later, the advent of distributed systems consisting of workstations and shared le servers resulted in much research on locality characteristics and their impact on caching at client [3] and le server caches [26]. Many recent studies have focussed on the characteristics of Web trac at clients [5] proxies [16, 28] and servers [1, 2, 6] Almeida et al. 1] used the LRUSM model to measure temporal locality in Web server access logs. Cao et al. 9] analyzed document inter reference times to establish the ....

....set compared to the USask data set. A possible explanation for this behavior is that references to more persistent hot documents tend to get captured at the lower level caches. 7 4. 2 Short Term Measure of Temporal Locality The Least Recently Used Stack Model (LRUSM) 22] has been widely used [1, 2, 26] to measure temporal locality. The LRUSM is a stack based ordering of referenced objects, according to their recency of reference (i.e. most recently referenced item on top (position 1) and the least recently referenced item on the bottom) For each reference in the request stream, the stack is ....

D. Willick, D. Eager, R. Bunt, Disk Cache Replacement Policies for Network File Servers, Proceedings of the 13th International Conference on Distributed Computing Systems (ICDCS), Pittsburgh, PA, May 1993, pp. 2-11.

No context found.

D.L. Willick, D.L. Eager, and R.B. Bunt, "Disk Cache Replacement Policies for Network Fileservers," Proc. 13th Int'l Conf. Distributed Computing Systems, May 1993.

No context found.

D. L. Willick, D. L. Eager, and R. B. Bunt. Disk cache replacement policies for network fileservers. In Proceedings of the 13th International Conference on Distributed Computing Systems (ICDCS '93), pages 2--11, Pittsburgh, Pennsylvania, May 1993.

No context found.

Darryl L. Willick, Derek L. Eager, and Richard B. Bunt. Disk Cache Replacement Policies for Network Fileservers. In International Conference on Distributed Computing Systems (ICDCS '93), pages 2--11, Pittsburgh, Pennsylvania, May 1993.

No context found.

D. L. Willick, D. L. Eager, and R. B. Bunt. Disk Cache Replacement Policies for Network Fileservers. In ICDCS, May 1993.

No context found.

D.L. Willick, D.L. Eager, and R.B. Bunt, "Disk Cache Replacement Policies for Network Fileservers," Proc. 13th Int'l Conf. Distributed Computing Systems, May 1993.

No context found.

D. L. Willick, D. L. Eager, and R. B. Bunt. Disk cache replacement policies for network fileservers. In Proceedings of the 13th International Conference on Distributed Computing Systems, May 1993.

No context found.

Darryl L. Willick, Derek L. Eager, and Richard B. Bunt. Disk Cache Replacement Policies for Network Fileservers. In International Conference on Distributed Computing Systems (ICDCS '93), pages 2--11, Pittsburgh, Pennsylvania, May 1993.

No context found.

D. L. Willick, D. L. Eager, and R. B. Bunt. Disk cache replacement policies for network fileservers. In Proceedings of the 13th International Conference on Distributed Computing Systems (ICDCS '93), pages 2--11, Pittsburgh, Pennsylvania, May 1993.

No context found.

D. Willick, D. Eager, and R. Bunt, "Disk Cache Replacement Policies for Network Fileservers", Proceedings of the 13th International Conference on Distributed Computing Systems (ICDCS), pp. 2-11, Pittsburgh, PA, May 1993.

No context found.

D. Willick, D. Eager, and R. Bunt, \Disk Cache Replacement Policies for Network File Servers," in Proceedings of the 13th International Conference on Distributed Computing Systems (ICDCS), Pittsburgh, PA, pp. 2-11, May, 1993.

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