A Static Analysis-based Cross-Architecture Performance Prediction Using\n Machine Learning

2019/06/18 by Newsha Ardalani, Urmish Thakker, Ardalani, Newsha +5 · 1 citation
Computer Science · #Software System Performance and Reliability #Parallel Computing and Optimization Techniques #Cloud Computing and Resource Management

paper · pdf · doi:10.48550/arxiv.1906.07840

Abstract

Porting code from CPU to GPU is costly and time-consuming; Unless much time\nis invested in development and optimization, it is not obvious, a priori, how\nmuch speed-up is achievable or how much room is left for improvement. Knowing\nthe potential speed-up a priori can be very useful: It can save hundreds of\nengineering hours, help programmers with prioritization and algorithm\nselection. We aim to address this problem using machine learning in a\nsupervised setting, using solely the single-threaded source code of the\nprogram, without having to run or profile the code. We propose a static\nanalysis-based cross-architecture performance prediction framework (Static\nXAPP) which relies solely on program properties collected using static analysis\nof the CPU source code and predicts whether the potential speed-up is above or\nbelow a given threshold. We offer preliminary results that show we can achieve\n94% accuracy in binary classification, in average, across different thresholds\n

Cited by

• The Tribes of Machine Learning and the Realm of Computer Architecture

Related

• Can Large Language Models Predict Parallel Code Performance?
• BB-ML: Basic Block Performance Prediction using Machine Learning Techniques
• Machine Learning Enabled Scalable Performance Prediction of Scientific\n Codes
• Towards CPU Performance Prediction: New Challenge Benchmark Dataset and Novel Approach
• Towards Universal Performance Modeling for Machine Learning Training on Multi-GPU Platforms