Enhancing the Performance of Condensed Graph’s Computation in Distributed Systems by Using Numerical Libraries as Super Nodes Samira Pakdel and Anne C. Elster Norwegian University of Science and Technology Sem Sælands vei 9, 7491 Trondheim, Norway samirap@idi.ntnu.no ABSTRACT Condensed Graph is a graph based programming model which is mainly used for the modeling of imperative, eager and lazy computation and also, it is a simple way to present a workflow. The biggest advantage of CG is to detect all independent instructions (nodes) for assigning them to the different computation devices in distributed and heterogeneous environments. This paper enhances the execution performance of condensed graphs in distributed systems by introducing super nodes for condensed graphs. Super nodes consist of a group of instructions treated as a single and atomic instruction. Accordingly, super nodes are more balanced in terms of computation time (T communication ) and communication time (T computation ). Therefore, the communication and computation time of the nodes can more likely happen at the same time (T overlap ) and ultimately reduces the overall graph execution time (T total ). In this paper, numerical libraries are used to create the super nodes. The results confirm that using super nodes with numerical libraries can have a huge impact on performance improvement in distributed systems. KEYWORDS Distributed computing, Condensed graph, Numerical libraries, Super nodes, Communication optimization. 1 INTRODUCTION Heterogeneous computing, distributed computing and parallel computing have been introduced for several decades and they have convergence in several aspects [1]. In fact, the most important goal for all of them is maximizing efficiency of computation power and decreasing the execution time [2]. Additionally, parallel computing known as a key to enhance the performance of computing. The main principle in parallel computing is that, all independent and unrelated instructions can be executed at the same time on different computation devices or different computing units of the same device (e.g. Multi core CPUs and GPUs). In this regard, graph implementation of an algorithm is one of the ways for detecting independent instructions (nodes) for executing in parallel. Condensed Graph (CG) is one of the successful graph models for computation modelling also, it can naturally be used in distributed, heterogeneous and parallel systems [3]. Therefore, in this study CG is chosen for further improvements and optimization. Additionally, optimizing and enhancing the performance of a computational algorithm is a critical task, affected by many factors. Numerical libraries can provide efficient and optimized computing implementations for Proceedings of the International Conference on Computer Science, Computer Engineering, and Social Media, Thessaloniki, Greece, 2014 ISBN: 978-1-941968-04-8 ©2014 SDIWC 86