Architectural Design of Grand Clos Collective Network for Supercomputers PLAMENKA BOROVSKA, DESISLAVA IVANOVA Computer Systems Department, Technical University of Sofia, Sofia BULGARIA pborovska@tu-sofia.bg, d_ivanova@tu-sofia.bg http://cs.tu-sofia.bg/ Abstract: As modern research and engineering computing methodologies, the demands for high-speed computational resources are growing at a rapid rate. The communication performance of collective network is a crucial factor influencing the communication performance of supercomputers which form the prevailing parallel architecture of modern high-performance computer systems. In this paper we have proposed a modular high-speed switch architectural design and a new collective network design of “Grand Clos” topology to meet the demands of efficient and high-speed communications on supercomputers. The communication performance parameters such as network latency and throughput are evaluated on the basis of parallel simulation models using the framework OMNET++ (C++, MPI) and run on IBM HS22 Blade Center. Analysis and comparison of simulation results has been performed and presented considering the impact of the traffic pattern and the packet size on the communication efficiency of the collective network. Key-Words: Collective Networks, Supercomputers, Grad Clos Architecture, Network Topology Design, OMNET++, Communication Performance Evaluation 1. Introduction Switch and collective network architecture designs are significantly influenced by next generation high-performance computer systems and supercomputer technology. The path toward realizing next generation exascale and z-scale computer systems is increasingly dependent on building supercomputers with thousands of processors. The supercomputer architecture interconnects of collective topology is a crucial factor in determining the computer performance, [1-4]. Collective networks vary with respect to throughput, latency, scalability, and cost. Network performance determines supercomputer performance for many applications. Therefore, the initial choice of a collective network design will affect the usability and performance of supercomputers. Interconnection design of collective networks is composed of a set of shared switch nodes and channels, and the topology of the network refers to the arrangement of these nodes and channels. Selecting the collective network topology design is the first step in designing a network because the routing algorithm and flow-control method depend on the topology. Selecting a good topology is the most important job of fitting the requirements of the collective network design to the available supercomputer technology [1], [3]. The goal of this paper is to suggest generalized and modular high-speed switch architectural design and a relevant collective network design of Grand Clos topology utilizing cut-through routing and to evaluate its communication efficiency for building up supercomputers. Communication performance of a high-speed switch and Grand Clos topology designs are performed by means of network simulations using OMNET++ and run on IBM Blade Center, located at High-Performance and GRID Computing Laboratory, Computer Systems Department, Technical University of Sofia. 2. OMNeT++ Framework OMNeT ++ is essentially a set of software tools and libraries that supports the development of simulation models. Most often OMNeT++ is used to develop models of computer networks and protocols, but the product can be used for the preparation of various models. OMNeT++ represents simulation environment, including specific libraries (simulation framework and library). It is built up of individual components called modules. Its main purpose is to be used for Recent Advances in Computer Science ISBN: 978-960-474-311-7 146