high performance computing on graphics processing units: hgpu.org

ATLAS is one of two general purpose collision detectors within the Large Hadron Collider, detecting millions of events per second. One tool for the eventual analysis of this data is TMVA, the Toolkit for Multi-Variate Analysis. Comprising of a number of machine learning techniques, it supports physicists in classifying events. This project forms a feasibility study into the use of GPU (Graphics Processing Unit) devices to parallelise TMVA to determine whether such techniques might lead to future performance gains for the framework. In particular, the Multi-layer perceptron, a class of neural network, is ported to the GPU programming platform CUDA. Performance when training single networks is generally comparable to CPU performance but show promise for future improvement. However, this parallelism of the GPU also allows multiple networks to be trained simultaneously, and this is leveraged to show significant performance gains over undertaking such a task in serial. The challenges and potential for these results to be applied across the TMVA framework is then considered and discussed.