Previous work has shown that cache line sizes impact performance differently for different desktop programs – some programs work better with small line sizes, others with larger line sizes. Typical processors come with a line size that is a compromise, working best on the average for a variety of programs. We analyze the energy impact of different line sizes, for 19 embedded system benchmarks, and we show that tuning the line size to a particular program can reduce memory access energy by 50% in some examples. Our data argues strongly for the need for embedded microprocessors to have configurable line size caches, and for embedded system designers to put effort into choosing the best line size for their programs.
Chuanjun Zhang, Frank Vahid, Walid A. Najjar