Comparing Controlflow and Dataflow for Tensor Calculus: Speed, Power, Complexity, and MTBF

• Published in: High Performance Computing Vol. 11203; pp. 329 - 346
• Main Authors: Kotlar, Milos, Milutinovic, Veljko
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 01.01.2018; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Big data; Dataflow computing; Tensor calculus; Ultimate dataflow
• ISBN: 3030024644; 9783030024642
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-030-02465-9_22

This article introduces ten different tensor operations, their generalizations, as well as their implementations for a dataflow paradigm. Tensor operations could be utilized for addressing a number of big data problems in machine learning and computer vision, such as speech recognition, visual object recognition, data mining, deep learning, genomics, mind genomics, and applications in civil and geo engineering. As the big data applications are breaking the Exascale barrier, and also the Bronto scale barrier in a not so far future, the main challenge is finding a way to process such big quantities of data. This article sheds light on various dataflow implementations of tensor operations, mostly those used in machine learning. The iterative nature of tensor operations and a large amount of data makes them situable for the dataflow paradigm. All the dataflow implementations are analyzed comparatively with the related control-flow implementations, for speedup, complexity, power savings, and MTBF. The core contribution of this paper is a table that compare the two paradigms for various data set sizes, and in various conditions of interest. The results presented in this paper are made to be applicable both for the current dataflow paradigm implementations and for what we believe are the optimal future dataflow paradigm implementations, which we refer to as the Ultimate dataflow. This portability was made possible because the programming model of the current dataflow implementation is applicable also to the Ultimate dataflow. The major differences between the Ultimate dataflow and the current dataflow implementations are not in the programming model, but in the hardware structure and in the capabilities of the optimizing compiler. In order to show the differences between the Ultimate dataflow and the current dataflow implementations, and in order to show what to expect from the future dataflow paradigm implementations, this paper starts with an overview of Ultimate dataflow and its potentials.