Adaptive java optimisation using instance-based learning

Abstract

This paper describes a portable, machine learning-based approach to Java optimisation. This approach uses an instancebased learning scheme to select good transformations drawn from Pugh's Unified Transformation Framework [11]. This approach was implemented and applied to a number of numerical Java benchmarks on two platforms. Using this scheme, we are able to gain over 70% of the performance improvement found when using an exhaustive iterative search of the best compiler optimisations. Thus we have a scheme that gives a high level of portable performance without any excessive compilations.

The ParaPhrase Project: Parallel Patterns for Adaptive Heterogeneous Multicore Systems

“…However, the interplay of such optimisations combined with the complexity and variety of target platforms often renders static mapping approaches far less effective than the mappings that can easily be achieved manually. Even in the single-processor setting it has been shown that semi-static approaches such as iterative optimisations [47,54] can improve the effectiveness of the optimisation process. The ParaPhrase approach avoids these problems by exploiting a more dynamic approach that can adapt to changing system conditions, given a good initial placement as its starting point.…”

Iterative optimization in the polyhedral model

“…He developed the GAPS framework [87] which used a genetic algorithm to traverse a search space of affine schedules for automatic parallelization. In addition, Long and O'Boyle [80] considered a search space of transformation sequences represented in the UTF framework [63]. Both of these approaches suffer from under-constraining the search space by considering all possible schedules, including illegal ones.…”

“…Our contribution is to reconcile fine-grain control of transformation heuristics -as opposed to optimization flag or pass selection [109,3] -with the guaranteed legality of the transformed program -as opposed to filtering approaches [87,80,81] or always-correct transformations [107,71].…”

“…These primitives would be defined in an abstract way and not for a pair of statements: for instance one can have f use(R, S), tile(R, S) in the generated program, those would be abstracted as f use,tile. Such an approach has been tested in the larger context of multidimensional (including illegal) affine schedules by Long et al [80,82].…”

A Cost-Aware Parallel Workload Allocation Approach Based on Machine Learning Techniques

“…Logistic regression is used in [7] to derive a predictive model that selects suitable optimizations to apply to each method based on code features. Instance-based learning is used in [13] to select suitable loop transformations within an optimization space of various loop re-ordering transformations.…”