Causal profiling is a new approach to software profiling that tells developers which code is important for performance. Conventional profilers report where programs spend their time running, but optimizing long-running code may not improve program performance. Instead of simply observing a program, a causal profiler conducts performance experiments to predict the effect of speeding up many different parts of a program. During each experiment, a causal profiler uses virtual speedup to create the effect of optimizing part of the program, and progress points to measure any change in program performance as a result of the virtual speedup. A causal profile summarizes the results of many performance experiments, telling developers exactly where performance tuning would be worthwhile. Using COZ, a prototype causal profiler for Linux, we improve the performance of Memcached by 9%, SQLite by 25%, and several PARSEC applications by as much as 68%.
