A comparison of online superpage promotion mechanisms

The amount of data that a typical translation lookaside buffer (TLB) can map has not kept pace with the growth in cache sizes and application footprints. As a result, the cost of handling TLB misses limits the performance of an increasing number of applications. The use of superpages, multiple adjacent virtual memory pages that can be mapped with a single TLB entry, extends a TLB's reach without significantly increasing its size or cost. The difficulty of identifying of identifying what sets of pages should be promoted to superpages combined with the overhead of performing these promotions restricts superpage use almost exclusively to wired system data structures. Previous studies have shown that simple online policies that decide to create superpages dynamically can be effective in reducing TLB penalties. In this paper we analyze the performance of online superpage promotion for nine benchmarks on a simulated HP PA-RISC system running a BSD Unix kernel. We extend previous work in two ways. First, we study the impact of creating superpages dynamically by remapping pages at the memory controller instead of copying pages to make them contiguous. The use of such a hardware mechanism affects the choice between two previously described superpage promotion policies. Previous work has shown that an online approximation to a competitive policy is the best choice. Our results show that having hardware support makes a greedy policy perform equally well. Second, we use execution-driven simulation, whereas previous studies have used trace-driven simulation. Our results show that the differences in accuracy are significant, especially when studying complex interactions between operating and modern architectures.