Pei Cao. Application-controlled file caching and prefetching (thesis). Technical Report TR-522-96, Princeton University, Computer Science Department, May 1996.  Home/Search   Document Not in Database   Summary   Related Articles   Check

.... These features include: 1) resource reservation (i.e. admission control) algorithms to provide QoS guarantees [32] 2) support for client pull and serverpush architectures [27] 3) support for fixed size and variable size blocks [34] and (4) support for data type specific caching techniques [3, 9]. Integrating these diverse techniques into the file system is a key challenge, since it can significantly complicate the file system architecture. For simplicity and modularity, Symphony employs a two layer architecture. The lower layer of Symphony implements a set of data type independent ....

....utilization, an integrated file system must support multiple data type specific failure recovery techniques. Caching Techniques A caching policy such as LRU can significantly improve response times for textual requests. It is well known that LRU performs poorly for sequential data accesses [3], and hence, is inappropriate for continuous media. Policies such as Interval Caching [7] that are tailored for sequential data accesses are more desirable for continuous media, but are unsuitable for textual data. Since no single policy is suitable for all data types, an integrated file system ....

[Article contains additional citation context not shown here]

P. Cao. Application Controlled File Caching and Prefetching. PhD thesis, Princeton University, 1996.

....loading a block that is not present in the cache. Whenever a block must be ejected, the global lowest cost block is chosen. The LRU SP system with controlled aggressive prefetching (hereafter referred to simply as LRU SP) presented by Cao, Karlin, Felten and Li [CFKL95b, CFKL95a, CFL94a, CFL94b, Cao96] addresses the problem via a two level procedure. Buffers are allocated among processes by the kernel using the LRUSP procedure, which ejects a block from the process containing the globally least recently used block. This process may, however, suggest a more appropriate block to eject based on ....

....to prefetch data from disk and improve the file buffer cache hit ratio. One is the informed prefetching and caching system, TIP2, developed by Patterson, Gibson, et al. [PGG 95] The other is the controlled aggressive LRU SP system of Cao, Karlin, et al. [CFKL95b, CFKL95a, CFL94a, CFL94b, Cao96] There are two key issues in designing such a system. First is how to balance buffer usage between prefetching and caching for a single process access stream. Second is how to allocate buffers among competing processes, both hinting and non hinting. The study of [KTP 96] focuses on the ....

Pei Cao. Application-Controlled File Caching and Prefetching. PhD thesis, Princeton University, 1996.

....from the client pull paradigm employed by existing proxies to service conventional requests. Finally, continuous media accesses are predominantly sequential in nature. Cache replacement policies such as LRU employed by existing proxy servers are known to be ineffective for sequential accesses [2]. Several research groups have investigated the design of specialized continuous media proxies to alleviate some of the drawbacks of conventional web proxies. Most of these efforts have focused on designing mechanisms for handling continuous media requests and a large number of mechanisms such as ....

P. Cao. Application Controlled File Caching and Prefetching. PhD thesis, Princeton University, 1996. 1 The current version of QLinux is available from http://www.cs.umass.edu/lass/software/qlinux.

.... including database management (DBMS) 22] garbage collection [23] and multi media applications [24] Furthermore, certain optimisations are possible when applicationspecific knowledge can be brought to bear, including improved page replacement and prefetching [17] better buffer cache management [25], and light weight signal handling [26] All of these may be realised by providing user level control over some of the virtual memory system. Unfortunately, none of the above mentioned operating systems provide QoS in their memory management: ffl No Isolation: applications which fault repeatedly ....

P. Cao. Application-Controlled File Caching and Prefetching. PhD thesis, Princeton University, January 1996.

.... database management systems (DBMS) 25] garbage collection [26] and multi media applications [27] Furthermore, certain optimisations are possible when applicationspecific knowledge can be brought to bear, including improved page replacement and prefetching [18] better buffer cache management [28], and light weight signal handling [29] All of these may be realised by providing user level control over some of the virtual memory system. Unfortunately, none of the above mentioned operating systems provide QoS in their memory management: ffl No Isolation: applications which fault repeatedly ....

Pei Cao. Application-Controlled File Caching and Prefetching. PhD thesis, Princeton University, January 1996.

....CPU cycles only in a very coarse manner, which is typically based on static limits on total consumption. The development of more powerful policies to control the consumption of such resources has been the focus of complimentary research in application specific paging [27, 20, 24] and file caching [9], disk bandwidth allocation [46, 47] and TCP buffer management [39] 4.5 The resource container hierarchy Resource containers form a hierarchy. The resource usage of a child container is constrained by the scheduling parameters of its parent container. For example, if a parent container is ....

P. Cao. Application Controlled File Caching and Prefetching. PhD thesis, Princeton University, Jan. 1996.

....from the client pull paradigm employed by existing proxies to service conventional requests. Finally, continuous media accesses are predominantly sequential in nature. Cache replacement policies such as LRU employed by existing proxy servers are known to be ineffective for sequential accesses [2]. Several research groups have investigated the design of specialized continuous media proxies to alleviate some of the drawbacks of conventional web proxies. Most of these efforts have focussed on designing mechanisms for handling continuous media requests and several mechanisms such as prefix ....

P. Cao. Application Controlled File Caching and Prefetching. PhD thesis, Princeton University, 1996.

....been two classes of API s proposed to allow the application to provide information to the operating system about its upcoming accesses. The first is cache management advice of the form, Cache file foo.data using the MRU cache replacement scheme. This approach has been implemented by Cao et al. [Cao96, CFL94b]. The second approach, known as disclosure, has been taken by Patterson et al. and also by Cao et al. A disclosure is a hint about future accesses given in the language of the existing I O interface. Thus, in Unix a disclosure of future accesses might be a list of file descriptor and byte range ....

....disks of the disk array busy; the fact that TIP2 performs substantially better suggests that sequential readahead schemes, even extremely aggressive schemes like OSF1 s, are insufficient to reliably keep the disks busy fetching needed data. Another system presented by Pei Cao and collaborators [Cao96] has shown a reduction of up to 50 in overall execution time on a different suite of applications, also modified to give hints to an operating system designed to accept them. In this case, the system had a single disk and the improvement was due largely to cache management. Thus, independent ....

[Article contains additional citation context not shown here]

Pei Cao. Application-Controlled File Caching and Prefetching. PhD thesis, Princeton University, 1996. (pp 28, 29, 32, 33, 59, 65)

....combines the multi process advantages of cost benefit allocation with the single process advantages of dynamic, load aware prefetching as demonstrated by Forestall. Another strategy for resource allocation in an informed prefetching and caching system is LRU SP, presented by Cao et al. [CFL94a, Cao96] which extends traditional LRU replacement to a mechanism for selecting the buffer supplier forced to evict a cache block. It is straightforward to implement the Forestall algorithm with LRU SP. With TIPTOE and LRU SP Forestall we have comparable prefetching and caching components and ....

....disk is not great enough to win any more buffers. At that point, the prefetcher issues the batch of requests and that disk is no longer considered idle. The prefetcher continues to try to fill batches for the remaining idle disks. 3 LRU SP Resource Allocation Pei Cao s LRU SP [CFL94a, CFL94b, Cao96] algorithm is an alternative to TIP2 s cost benefit strategy for allocating buffers among multiple processes in the presence of hints disclosing future accesses. The goal of the algorithm is to adapt the time tested LRU algorithm s fairness and performance qualities to this new domain. LRU SP ....

[Article contains additional citation context not shown here]

Pei Cao. Application-Controlled File Caching and Prefetching. PhD thesis, Princeton University, 1996.

....LRU SP (LRU with Swapping and Placeholders) allocation policy is designed based on the above principles. It can be proven that, in a simplified theoretical model that describes the case of application controlled caching without prefetching and disk scheduling, LRU SP satisfies all three criteria [Cao96] LRU SP operates by maintaining an LRU list of all in cache blocks. The list is in the order of the blocks last references, except for two modifications (explained below) When a process misses in the cache or issues a prefetch, some block must be evicted from the cache in order to make room ....

....the system to detect when the application s choice is not as good as that of the default policy (e.g. B is accessed before A) in this case, the erroneous process will give up a block for the extra cache miss, instead of other processes giving up a block. More details can be found in [CFL94a, Cao96] Early experiments with application controlled file caching showed that LRUSP performs quite well, and the swapping and placeholders mechanisms are crucial to performance improvement in multi process environments [CFL94b] For example, if the allocation scheme does not use swapping but simply ....

Pei Cao. Application-Controlled File Caching and Prefetching. PhD thesis, Princeton University, January 1996. Also published as technical report CS TR-522-96.

....Let A 1 Gammak (t) be the first k elements of A(t) and let H be the operation that substitutes every a i (t)hh i (t)i by h i (t) and every hh j (t)i by h j (t) Then at every r t , 1 t n, H(A 1 Gammak (t) G(t) 1) The proof is essentially by case study and inductions. The details are in [1]. The techniques used in proving Lemma 1 also prove the following lemma. Lemma 2 At every reference r t of , let N v (P i ; Alg i ; t) be the number of virtual blocks in A(t) that are in P i s access sequence. Then for all i, FLRU GammaS P (P i ; Alg i ; t) N v (P i ; Alg i ; t) F ....

Pei Cao. Application-Controlled File Caching and Prefetching. PhD thesis, Princeton University, December 1995. Also published as technical report CS TR-522-95.

No context found.

Pei Cao. Application-controlled file caching and prefetching (thesis). Technical Report TR-522-96, Princeton University, Computer Science Department, May 1996.

No context found.

P. Cao. Application Controlled File Caching and Prefetching. PhD thesis, Princeton University, 1996.

No context found.

P. Cao, Application Controlled File Caching and Prefetching. PhD thesis, Princeton University, 1996.

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