While this work has shown that file reference patterns provide valuable information for caching, and use of such information can greatly reduce I/O latency, we have also found certain areas that require further study. We hope to examine the following issues.
The paging of predictive data to and from disk is critical to the success of predictive prefetching. While our implementation was done in a manner that facilitates such functionality, we have not directly addressed this issue.
The idea of partition jumping would use multiple partitions to continue in a sequence past the end of one partition and into another partition that begins with the last n symbols of the sequences. This would allow EPCM to make predictions deeper than the partition size. This new method would generate predictions with EPCM as before, but when a descendent with no children was found, the last n symbols in the pattern would be used as an n-order context into a new partition from which predictions would continue. This would enable EPCM to look into other partitions once it has reached the end of the current partition, and enable smaller partitions to predict further ahead than their partition size would normally allow.
In our test environment, we ran the same benchmark test consistently, so our models saw no variation. As a result, they generated no erroneous prefetching. It would be instructive to use trace-based simulations to investigate how often our models would incorrectly prefetch. If we then forced an implementation to make this percentage of incorrect prefetches, we could gauge the impact of incorrect prefetching on the system as a whole.