Chen, K., Write Caching in Distributed File Systems, M.Sc. thesis, Department of Computational Science, University of Saskatchewan, August 1994.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....differently. The objective is to maintain a minimum level of cache accuracy. By estimating the lifetime of a cached object and its age, the application could determine the degree of accuracy of the object in discussion. Hints that are highly accurate ensure good performance benefits. In [11], the issue of write caching in distributed systems is examined. Write policies used in traditional file system caches use either write through or periodic write back that may result in little benefit in general distributed settings. Here, systems with client and server non volatile caches are ....

K. Chen, R. Bunt, and D. Eager. Write Caching in Distributed File Systems. In Proccedings of the 15th IEEE International Conference on Distributed Computing Systems, pages 457--466, May-June 1995.

....system mutating operations back to a file server. In this way, the results of a write at a client are reflected in the file system rather than just stored in a log file. A cache s write back policy determines when and how write backs occur. There are performance advantages to delaying write backs [1, 5, 12]. For example, multiple mutating operations can be accommodated in a single write back. Examples of write back policies include write back on close and write back on reconnect. Write backs can be initiated either by the client [10, 17, 19] or by the server [48] In a weakly connected environment, ....

K. Chen. Write caching in distributed file systems. Master's thesis, University of Saskatchewan, Saskatoon, SK, 1994.

....this to represent typical file access activity within a university research environment. In the interests of conserving space while, at the same time, presenting comparable results, results from only a single trace are presented in this paper. Results from a wider set of traces can be found in [Chen 94] The trace selected came from tracing a set of simulation experiments used to generate the results in [Willick et al. 93] This particular trace was chosen because it involved a good deal of file activity (738,489 read events and 8,547 write events over a span of 3.23 hours) Each time stamped ....

....of LRU on the client (to take advantage of the temporal locality in the reference stream at the client) and WBT on the server (to amortize I O costs) appears to be the best combination. 4 Other Approaches Several other approaches to writing back cached data are possible, and are considered in [Chen 94] Two of these are reviewed briefly in this section. 4.1 Piggybacking Since read requests that miss at both the client cache and the server cache necessitate immediate disk access, it makes sense to take out any writes that might be in the same area of the disk at the same time. By ....

Chen, K., Write Caching in Distributed File Systems, M.Sc. thesis, Department of Computational Science, University of Saskatchewan, August 1994.

....are later negated by subsequent updates. For example, a file may be created, written to, then deleted, removing any need for previous updates to the file to be sent to a file server. This means that the longer write backs are delayed, the fewer write backs will actually need to be performed [3]. Further complicating the write back issue is the fact that in a limited bandwidth environment, it may not be desirable to consume bandwidth with write backs if this has too great of a negative impact on the performance of other (e.g. read request and replies) network usage. For all of the ....

Chen, K., Bunt, R., and Eager, D. Write caching in distributed file systems. In Proc. Thirteenth International Conference on Distributed Computing Systems (Vancouver, BC, May 1995), pp. 457--466.

....application (a write hit) the results of this change must always find their way back to the disk to maintain data consistency, and so some disk activity is required whether or not the requested block is in the cache. This presents further cache management challenges, some of which are explored in [4]. A distributed client server environment introduces additional factors that also bear on performance. First, a network is interposed between the client workstation and the file server, thereby introducing yet another opportunity for delay. A second, and potentially more serious, complication is ....

....writes that must go through to the disk whether they were hits or not. This filtering of the references could have a profound impact on the performance of a localitybased cache management strategy at the server. This is confirmed by the results of performance experiments reported elsewhere ( 3] [4]) Figure 5 shows the effect of client caching on the reference stream seen at the server. These plots of LRU stack depth and LFU stack depth were made after the effects of client caching to show the reference behaviour in the filtered reference streams. With a client cache size of 0 (i.e. no ....

Chen, K., Bunt, R.B. and Eager, D.L., "Write Caching in Distributed File Systems", Proc. 15th International Conference on Distributed Computing Systems, Vancouver, Canada, May 1995, 457466.

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