Skip to Main Content
Recent technology trends are leading to the possibility of computer systems having last level caches significantly larger than those that exist today. Traditionally, cache effectiveness has been modeled through trace-driven simulation tools, however, these tools are not up to the task of modeling very large caches. Because of the limited length of available traces, the tools cannot capture behavior across long enough periods of time to adequately simulate a very large cache. We present mprofiler, a tool that characterizes the memory access pattern of workloads, and present a novel hybrid modeling technique that models cache behavior across a much larger time scale than previously possible. Our methodology combines memory access patterns captured by different tools (e.g., mprofiler) at different time scales and develops analytical techniques that allow spanning the required time frame and predicting the performance of very large caches.