Microarchitecture Sensitive Empirical Models for Compiler Optimizations

Abstract

This paper proposes the use of empirical modeling techniques for building microarchitecture sensitive models for compiler optimizations. The models we build relate program performance to settings of compiler optimization flags, associated heuristics and key microarchitectural parameters. Unlike traditional analytical modeling methods, this relationship is learned entirely from data obtained by measuring performance at a small number of carefully selected compiler/microarchitecture configurations. We evaluate three different learning techniques in this context viz. linear regression, adaptive regression splines and radial basis function networks. We use the generated models to a) predict program performance at arbitrary compiler/microarchitecture configurations, b) quantify the significance of complex interactions between optimizations and the microarchitecture, and c) efficiently search for 'optimal' settings of optimization flags and heuristics for any given microarchitectural configuration.Our evaluation using benchmarks from the SPEC CPU2000 suits suggests that accurate models (< 5% average error in prediction) can be generated using a reasonable number of simulations. We also find that using compiler settings prescribed by a model-based search can improve program performance by as much as 19% (with an average of 9.5%) over highly optimized binaries.

Program optimization space pruning for a multithreaded gpu

“…Work has also been done to study the interaction among different optimizations and between optimizations and the hardware without a full search. These range from those based on analytical models [9,17] to those that use statistical models [13] and those that utilize adaptive learning and intelligent search techniques [3,4,26,27] to find an optimal configuration. Finally, work by the SPIRAL project [2] generally uses an iterative approach to find desirable code, whereas we do not.…”

Measuring empirical computational complexity

“…Vaswani et al [VTSJ07] build regression models that relate a benchmark's performance to micro-architectural parameters, compiler optimization flags, and associated compiler optimization heuristic parameters (for instance maximum loop unrolling). They use these models to (a) predict performance at arbitrary compiler and micro-architecture settings, (b) identify micro-architectural features that interact (both beneficially and detrimentally) with compiler optimization settings, and finally (c) find optimal settings for a particular program.…”

Microarchitectural Design Space Exploration Using an Architecture-Centric Approach

“…This is similar to the character- isation of the space that we conduct in section 7. The other schemes are similar in terms of accuracy [18,27]. However, none of these papers characterise the space that they are exploring by showing the correlation of microarchitectural parameters to the best configurations.…”

“…• Whenever a new program is considered, a new predictor must be trained and built, meaning there is a large overhead even if the designer just wants to compile with a different optimisation level [27]. Our approach learns across programs and captures the behaviour of the architecture rather than the program itself;…”