Static Program Features Research Articles

Predicting Loop Vectorization through Machine Learning Algorithms

Automatic vectorization is often utilized to improve the speed of compute-intensive programs on current CPUs. However, there is enormous space for improvement in present compiler auto-vectorization capabilities. Execution with optimizing code on these resource-controlled strategies is essential for both energy and performance efficiency. While vectorization suggests major performance developments, conventional compiler auto-vectorization techniques often fail. This study investigated the prospective of machine learning algorithms to enhance vectorization. The study proposes an ensemble learning method by employing Random Forest (RF), Feedforward Neural Network (FNN), and Support Vector Machine (SVM) algorithms to estimate the effectiveness of vectorization over Trimaran Single-Value Code (TSVC) loops. Unlike existing methods that depend on static program features, we leverage dynamic features removed from hardware counter-events to build efficient and robust machine learning models. Our approach aims to improve the performance of e-business microcontroller platforms while identifying profitable vectorization opportunities. We assess our method using a benchmark group of 155 loops with two commonly used compilers (GCC and Clang). The results demonstrated high accuracy in predicting vectorization benefits in e-business applications.

A Framework for OpenCL Task Scheduling on Heterogeneous Multicores

We present an intelligent scheduling framework which takes as input a set of OpenCL kernels and distributes the workload across multiple CPUs and GPUs in a heterogeneous multicore platform. The framework relies on a Machine Learning (ML) based frontend that analyzes static program features of OpenCL kernels and predicts the ratio in which kernels are to be distributed across CPUs and GPUs. The framework provides such static analysis information along with system state information like runtime availability details of computing cores using well defined programming interfaces. Such interfaces are to be utilized by a user specified scheduling strategy. Given such a scheduling strategy, the framework generates device specific binaries and dispatches them across multiple devices in the heterogeneous platform as per the strategy. We test our scheduling framework extensively using different OpenCL task mixes of varying sizes and computational nature. Along with the scheduling framework, we propose a set of novel partition-aware scheduling strategies for heterogeneous multicores. Our proposed approach yields considerably better results in terms of schedule makespan when compared with the current state of the art ML based methods for scheduling of OpenCL workloads across heterogeneous multicores.

Predicting power consumption of GPUs with fuzzy wavelet neural networks

Prediction and optimization of power consumption have become an essential issue in the field of General-purpose computing on graphic processing units (GPUs) because of the increasing prevalence of GPUs and the constraints of energy consumption. However, previous approaches to build power models need to extract program features related to power consumption from the performance counter or GPU emulators. These approaches are unable to estimate power consumption of applications for software designers owing to the lack of information about detailed GPUs architecture or the supporting of emulators. In this study, we explore a novel method to model GPU power consumption of applications in computing process. By using program slicing, we decompose the source code of applications into slices and extract the power-related static program features. The slicing is used as a basic unit to train a power model based on fuzzy wavelet artificial neural networks. This step allows programmers to investigate the power profile of their applications and identify the code areas with higher energy consumption. To improve prediction accuracy, we further divide the GPUs applications by the branch structure into two categories: sparseness-branch and denseness-branch. The power model is proposed for sparseness-branch programs based on slicing. For denseness-branch programs, probabilistic slicing is utilized to reduce the invalid slices in order to improve accuracy. The models are empirically validated by using typical GPU benchmarks and the results are compared with the measured power. Overall, the average error of our power models is less than 6%.

Automatic problem size sensitive task partitioning on heterogeneous parallel systems

In this paper we propose a novel approach which automatizes task partitioning in heterogeneous systems. Our framework is based on the Insieme Compiler and Runtime infrastructure. The compiler translates a single-device OpenCL program into a multi-device OpenCL program. The runtime system then performs dynamic task partitioning based on an offline-generated prediction model. In order to derive the prediction model, we use a machine learning approach that incorporates static program features as well as dynamic, input sensitive features. Our approach has been evaluated over a suite of 23 programs and achieves performance improvements compared to an execution of the benchmarks on a single CPU and a single GPU only.

Feature Subset Selection for Hot Method Prediction using Genetic Algorithm wrapped with Support Vector Machines

Problem statement: All compilers have simple profiling-based heuristics to identify and predict program hot methods and also to make optimization decisions. The major challenge in the profile-based optimization is addressing the problem of overhead. The aim of this work is to perform feature subset selection using Genetic Algorithms (GA) to improve and refine the machine learnt static hot method predictive technique and to compare the performance of the new models against the simple heuristics. Approach: The relevant features for training the predictive models are extracted from an initial set of randomly selected ninety static program features, with the help of the GA wrapped with the predictive model using the Support Vector Machine (SVM), a Machine Learning (ML) algorithm. Results: The GA-generated feature subsets containing thirty and twenty nine features respectively for the two predictive models when tested on MiBench predict Long Running Hot Methods (LRHM) and frequently called hot methods (FCHM) with the respective accuracies of 71% and 80% achieving an increase of 19% and 22%. Further, inlining of the predicted LRHM and FCHM improve the program performance by 3% and 5% as against 4% and 6% with Low Level Virtual Machines (LLVM) default heuristics. When intra-procedural optimizations (IPO) are performed on the predicted hot methods, this system offers a performance improvement of 5% and 4% as against 0% and 3% by LLVM default heuristics on LRHM and FCHM respectively. However, we observe an improvement of 36% in certain individual programs. Conclusion: Overall, the results indicate that the GA wrapped with SVM derived feature reduction improves the hot method prediction accuracy and that the technique of hot method prediction based optimization is potentially useful in selective optimization.