The Implementation of File Migration 53 5 . 3 Evaluation of Migration Strategies 55 5 . 4 Stack Algorithms and Stack Processing

This thesis considers the problem of automatic file migration. In a computer system, program and data files are stored on secondary memory such as discs. If the capacity of the secondary memory is not sufficient to hold all the files, some sort of backup memory (e .g. magnetic tape) is used to supplement secondary storage. The transfer of files between secondary and backup storage is refered to here as "migration ." Migration may be done explicitly by each user or automatically by the operating system. This work studies automatic migration strategies. We seek the migration strategy which maximizes the probability that a file that is accessed will be found in secondary memory and which minimizes the migration activity. The problem of deciding .which files should m1grate, the file replacement problem, is analogous to the well-studied page replacement problem. It reduces to the problem of predicting the next inter-access interval for a file. An important difference between the file replacement problem and page replacement is that files are not equal in size. In addition, the size and time scales for file accesses are much larger than for page references. Consequently most page replacement results need to be reconsidered. Trace data of file system activity on a general purpose computer system have been collected for one year. This type of data has not been published previously. The data provide the basis for a quantitative study of migration. Statistical analysis of the trace data reveal characteristics of files stem activity. Special attention is paid to possible correlations of files' inter-access intervals to other factors, which may be useful in predicting inter-access intervals. Various possible file replacement, or migration, strategies are presented. Two non-equivalent criteria, the miss ratio and migration traffic rate, are used for evaluating strategies. LRU (least recently used) and MIN (the theoretically optimum strategy for fixed sized page)s are studied. Other strategies, including LFU (least frequently used), that are suggested by the results of statistical analysis, are examined. The fact that files are not equal in size can be used to advantage and the previously mentioned replacement strategies are shown to be improved when the size of a file is taken into consideration. Finally some non-realizable strategies using knowledge of files' future accesses are presented for comparison with the practical strategies already mentioned. A simulation model of file system activity is examined. The model is useful for generating artificial trace data. …