The following items are intended as areas of future exploration:
Trie memory requirements -- Our current implementation was designed as a proof of concept without concern for the memory usage of our predictive trie. Both Griffioen [5] and Curewitz [4] successfully address this issue. We hope to expand on their work, in conjunction with data compression techniques. We expect that our predictive cache will see significantly improved efficiency after we implement a fixed limit on the number of children each node can have. Additionally, methods that give more weight to more recently seen patterns and purge old informations not only reduce the model size, but also can improve performance.
Predicting from grandparents -- It is quite possible for one file open event to be the cause of two future events that occur closely. As a result, the order of these two events may vary. For example, an open of file B causes opens to files A and C, so we would have one of two resulting sequences BAC or BCA. We intend to investigate using all the descendants of a context to predict possible variations in the ordering of events. Using the final tree from Figure 1, if we see an open of file B then we would not only prefetch file C but also file A as well. Such a forward-looking prediction would enable us to avoid cache misses for the sequence BAC as well as BCA.
Modifications to the prefetching algorithm -- Curewitz's [4] approach to prefetching is to select the n most likely items (where n is a parameter of the model). We intend to investigate a combination of this method and the threshold based selection that we have used by placing a limit on the number of files that can be prefetched. We also intend to investigate the effect of having different threshold settings for each order of the model.
Predictive replacement -- While our cache is based on a predictive model to prefetch files, it still uses LRU to determine which files to expel from the cache when space is needed. Using our predictive model to determine cache replacements may offer further improvements in performance.
Read wait times -- While cache hit ratios are commonly used to measure the performance of a caching algorithm, we are mostly concerned with the amount of time that is spent waiting for file access events to complete. Our intent is to extend our simulation to include read-wait times allowing us to account for the additional I/O load generated by prefetching.
Selection by cost comparison -- Finally we intend to explore a prefetching selection model that uses the estimated probabilities of our model, in conjunction with other factors, such as memory pressure and file size, to estimate the cost, in terms of read-wait times. These estimates would be used in each case to decide if it was more beneficial to prefetch or not. Our hope is that such a parameterless selection method will adjust to changing file system behavior.