Limits of parallelism using dynamic dependency graphs

The advance of multi-core processors has led to renewed interest in extracting parallelism from programs. It is sometimes useful to know how much parallelism is exploitable in the limit for general programs, to put into perspective the speedups of various parallelisation techniques. Wall's study [19] was one of the first to examine limits of parallelism in detail. We present an extension of Wall's analysis of limits of parallelism, by constructing Dynamic Dependency Graphs from execution traces of a number of benchmark programs, allowing us better visualisation of the types of dependencies which limit parallelism, as well as flexibility in transforming graphs when exploring possible optimisations. Some of the results of Wall and subsequent studies are confirmed, including the fact that average available parallelism is often above 100, but requires effective measures to resolve control dependencies, as well as memory renaming. We also study how certain compiler artifacts affect the limits of parallelism. In particular we show that the use of a spaghetti stack, as a technique to implicitly rename stack memory and break chains on true dependencies on the stack pointer, can lead to a doubling of potential parallelism.