Autotuning algorithmic choice for input sensitivity

Abstract

A daunting challenge faced by program performance autotuning is input sensitivity, where the best autotuned configuration may vary with different input sets. This paper presents a novel two-level input learning algorithm to tackle the challenge for an important class of autotuning problems, algorithmic autotuning. The new approach uses a two-level input clustering method to automatically refine input grouping, feature selection, and classifier construction. Its design solves a series of open issues that are particularly essential to algorithmic autotuning, including the enormous optimization space, complex influence by deep input features, high cost in feature extraction, and variable accuracy of algorithmic choices. Experimental results show that the new solution yields up to a 3x speedup over using a single configuration for all inputs, and a 34x speedup over a traditional one-level method for addressing input sensitivity in program optimizations.

Input responsiveness: using canary inputs to dynamically steer approximation

“…• Algorithm Choice [3,15] we use IRA to choose between five different algorithmic implementations of Jmeint that offer different accuracy-performance tradeoffs in computing whether pairs of 3D triangles intersect. The most complex algorithm is the exact algorithm, while the simplest algorithm uses computationally cheap heuristics that work well only when triangles are far apart.…”

“…The general purpose approximate computing techniques typically applied to regularly-structured computations, such as loop perforation [27,45], algorithm selection [3,15], and numerical approximation [25], have been important and successful vehicles for realizing approximation in practice. These approaches can be realized on commodity hardware, apply to a variety of problem types, and are straightforward for programmers to implement.…”

“…• Effective, Automatic Approximationwe evaluate IRA on real server hardware, demonstrating that it produces effective input responsive approximations. The evaluation incorporates four well-known classes of approximate techniques from the literature: loop perforation [27,45], tiling [50], algorithm choice [3,15] and numerical approximation [25] within 13 image processing, machine learning data mining and computer vision applications ( §5).…”

End-to-End Deep Learning of Optimization Heuristics

“…We provide these inputs as a means of increasing the flexibility of our system, for example, to support applications in which the optimization heuristic depends on dynamic values which cannot be statically determined from the program code [3,24]. When present, the values of auxiliary inputs are concatenated with the output of the language model, and fed into the heuristic model.…”

“…Machine learning has emerged as a viable means in automatically constructing heuristics for code optimization [3,4,24,[37][38][39]. Its great advantage is that it can adapt to changing hardware platforms as it has no a priori assumptions about their behavior.…”