When an algorithm is processing some data, e.g. numbers are sorted, an image is filters or a neural network is trained, the foremost objective has been to do as this fast as possible. For decades new processors have become faster, by executing more operations per second, and new algorithms have been more efficient with shorter runtimes. Today, however, computers, from small devices to large data centres and supercomputers, consume a significant and growing amount of electrical energy. In order to achieve more sustainable computing, we need to reduce the energy consumption of computation. While this is being addressed on a technological level, with smaller processors, memories and networks, the algorithmic side receives little attention.
In this project you will investigate and compare computing tasks, for example processing a batch of images or training a neural network, in terms of the energy that is needed to complete the computation. For many problems there exist alternative approaches, only think of the many ways numbers can be sorted. We want to find out which of these alternatives for a given problem needs less energy to obtain the result. This will be done by studying the literature on the topic, identifying computational problems and algorithms with alternative approaches and then measuring and comparing their energy consumption of a given lab computer of the PARC (Parallel and Reconfigurable Computing) lab. Energy consumption can vary not only by how long it takes to compute, but by how many processors are involved, how much data is transferred and how often processors are idle. We aim to characterise patterns and approaches that can be used to identify the best techniques to be used from areas like data science and machine learning which have strong saving potential and a growing appetite for computing power.
Undergraduate
None
Parallel Computing (405.956, Lab)