Causal Profiling (Full Presentation)

Developers use profilers to identify code that should be op-timized. For each block of code, developers would like to know “How much will optimizing this code improve perfor-mance? ” This is a question of causality, which traditional profilers cannot answer. Traditional profilers only observe execution. They measure the time spent in each function and the number of times each function was called. For serial pro-grams, these measurements are sufficient to infer causality; however, concurrent and asynchronous programs present a problem. Threads in a parallel program can interfere, and not all code contributes to a program’s execution time. Causal Profiling We introduce Causal Profiling, a novel profiling technique that relies on controlled experimentation rather than passive observation. Profilers that simply observe a program’s exe