Automatic Parallelization of Affine Loops using Dependence and Cache analysis in a Binary Rewriter

Thumbnail Image


Publication or External Link






Today, nearly all general-purpose computers are parallel, but nearly all software running on them is serial. Bridging this disconnect by manually rewriting source code in parallel is prohibitively expensive. Automatic parallelization technology is therefore an attractive alternative.

We present a method to perform automatic parallelization in a binary rewriter. The input to the binary rewriter is the serial binary executable program and the output is a parallel binary executable. The advantages of parallelization in a binary rewriter versus a compiler include (i) compatibility with all compilers and languages; (ii) high economic feasibility from avoiding repeated compiler implementation; (iii) applicability to legacy binaries; and (iv) applicability to assembly-language programs.

Adapting existing parallelizing compiler methods that work on source code to work on binary programs instead is a significant challenge. This is primarily because symbolic and array index information used in existing compiler parallelizers is not available in a binary. We show how to adapt existing parallelization methods to achieve equivalent parallelization from a binary without such information. We have also designed a affine cache reuse model that works inside a binary rewriter building on the parallelization techniques. It quantifies cache reuse in terms of the number of cache lines that will be required when a loop dimension is considered for the innermost position in a loop nest. This cache metric can be used to reason about affine code that results when affine code is transformed using affine transformations. Hence, it can be used to evaluate candidate transformation sequences to improve run-time directly from a binary.

Results using our x86 binary rewriter called SecondWrite on a suite of dense- matrix regular programs from Polybench suite of benchmarks shows an geomean speedup of 6.81X from binary and 8.9X from source with 8 threads compared to the input serial binary on a x86 Xeon E5530 machine; and 8.31X from binary and 9.86X from source with 24 threads compared to the input serial binary on a x86 E7450 machine. Such regular loops are an important component of scientific and multi- media workloads, and are even present to a limited extent in otherwise non-regular programs.

Further in this thesis we present a novel algorithm that enhances the past techniques significantly for loops with unknown loop bounds by guessing the loop bounds using only the memory expressions present in a loop. It then inserts run-time checks to see if these guesses were indeed correct and if correct executes the parallel version of the loop, else the serial version executes. These techniques are applied to the large affine benchmarks in SPEC2006 and OMP2001 and unlike previous

methods the speedups from binary are as good as from source. We also present results on the number of loops parallelized directly from a binary with and without this algorithm. Among the 8 affine benchmarks among these suites, the best existing binary parallelization method achieves an geo-mean speedup of 1.33X, whereas our method achieves a speedup of 2.96X. This is close to the speedup from source code of 2.8X.