Automatic Parallelization of Sequential Programs
Authors:
Peter Kraft,
Amos Waterland,
Daniel Y Fu,
Anitha Gollamudi,
Shai Szulanski,
Margo Seltzer
Abstract:
Prior work on Automatically Scalable Computation (ASC) suggests that it is possible to parallelize sequential computation by building a model of whole-program execution, using that model to predict future computations, and then speculatively executing those future computations. Although that prior work demonstrated scaling, it did not demonstrate speedup, because it ran entirely in emulation. We t…
▽ More
Prior work on Automatically Scalable Computation (ASC) suggests that it is possible to parallelize sequential computation by building a model of whole-program execution, using that model to predict future computations, and then speculatively executing those future computations. Although that prior work demonstrated scaling, it did not demonstrate speedup, because it ran entirely in emulation. We took this as a challenge to construct a hardware prototype that embodies the ideas of ASC, but works on a broader range of programs and runs natively on hardware. The resulting system is similar in spirit to the original work, but differs in practically every respect.
We present an implementation of the ASC architecture that runs natively on x86 hardware and achieves near-linear speedup up to 44-cores (the size of our test platform) for several classes of programs, such as computational kernels, map-style programs, and matrix operations. We observe that programs are either completely predictable, achieving near-perfect predictive accuracy, or totally unpredictable, and therefore not amenable to scaling via ASC-like techniques. We also find that in most cases, speedup is limited only by implementation details: the overhead of our dependency tracking infrastructure and the manipulation of large state spaces. We are able to automatically parallelize programs with linked data structures that are not amenable to other forms of automatic parallelization.
△ Less
Submitted 29 July, 2018;
originally announced September 2018.
