Visualizing cache effects on I/O workload predictability

We describe our experience graphically visualizing data access behavior, with a specific emphasis on visualizing the predictability of such accesses and the consistency of these observations at the block level. Such workloads are more frequently encountered after filtering through intervening cache levels and in this paper we demonstrate how such filtered workloads pose a problem for traditional caching schemes. We demonstrate how prior results are consistent across both file and disk access workloads. We also demonstrate how an aggregating cache based on predictive grouping can overcome such filtering effects. Our visualization tool provides an illustration of how file workloads remain predictable in the presence of intervening caches, explaining how the aggregating cache can remain effective under what would normally be considered adverse conditions. We further demonstrate how the same predictability remains true with physical block workloads.