Performance analysis of cache oblivious Algorithms in the Fresh Breeze memory model

Abstract

The Fresh Breeze program execution model was designed for easy, reliable and massively scalable parallel performance. The model achieves these goals by combining a radical memory model with efficient fine-grain parallelsim and managing both in hardware. This presents a unique opportunity for studying program execution in a system whose memory behavior is not well understood. In this thesis, I studied the behavior of cache-oblivious algorithms within the Fresh Breeze model by designing and implementing a cache-oblivious matrix multiply within the Fresh Breeze programming framework, as well as a cache-naive algorithm for comparison. The algorithms were implemented in C, using the Fresh Breeze run-time libraries, and profiled on a simulated Fresh Breeze processor. I profiled both programs across a range of problem sizes, memory speeds and memory types in order to best understand their behavior and accurately characterize their performance.