ILLIAC 6

Overview

The convergence of computing and communications has profound implications for supercomputing that have yet to be fully realized in a single system that is at once a very high-performance communications platform and a very high-performance computational platform. It is the goal of the ILLIAC 6 project to create such a system.

The consequences of this convergence for high-performance computer and communication systems can be best seen by considering the computational aspects of high-performance communications systems on the one hand, and the communication requirements of modern supercomputers on the other.

Consider a high-end network router. It sits at the junction of a number of high-speed networks and examines the packets on those networks in order to forward them to their destination, filter them, prioritize them, convert them between formats and protocols, and so on. To the extent that the packets must be examined deeply in order to perform these functions, the router is performing a massive computation requiring exorbitant memory bandwidth and computational bandwidth -- and it is doing so in real time.

The ability to process packets based on their contents suggests a number of compelling applications that are infeasible today. For example, a firewall that sits between an enterprise or government network and the internet at large might wish to examine network data deeply to discover the escape of sensitive material from the organization, or the entrance of suspicious information into the organization. The firewall is the natural place to perform such security functions, for the simple reason that information flows in both directions pass though it. Once sensitive information has escaped the organization, no amount of computation can recover it. This fact makes it compelling to have very high computational capabilities in communication gear that is used in whole or in part for security functions.

At the same time, a class of modern supercomputing applications has emerged in which huge volumes of data arrive or are produced in real-time over high-bandwidth communication channels, and significant computations must be performed on them, also in real time. In these applications the volume of data is such that storage for off-line processing is impractical, and is moot besides as the data is being constantly generated from live sources such as sensor networks -- for example, arrays of video cameras, microphones, or telescopes -- or from a communication channel itself, as in the case of advanced security operations performed on a high-speed network or wireless channel. The computation to be performed on the data in these applications may include filtering (reducing the raw data to "interesting" features), computer vision or speech recognition, decryption (e.g., code breaking), and so on. Many of these computations are complex and difficult even when performed offline on a single data set. They become monumentally difficult supercomputing challenges when they are to be performed in real time on streams of live data.

The ILLIAC 6 project aims to design, construct, and deploy a system that is at once a massively parallel supercomputer and a high-performance communication system. A modern communications processor -- awe have chosen the Analog Devices Incorporated TigerSHARC® -- is the central component of the machine. The system is being designed to address modern real-time supercomputing challenges in the area of security and live processing of media and scientific data on wireless and wireline high-speed communication channels, as described briefly above.

The TigerSHARC combines high rates of computation, low power dissipation, excellent interprocessor communication, sizeable on-chip memory, native handling of the data types common to media processing and communications, and very good reliability. ILLIAC 6 is designed to achieve record-breaking performance on modern supercomputing applications in the areas of security, cryptography, stream processing, sensor network filtering, packet processing, computer vision, discrete optimization, and so on. It will be attached directly to optical networks which will act as its primary input/output medium. The system is designed to be simple, scalable, reliable, easy to program, and cost-effective.

The ILLIAC 6 project encompasses all aspects of the system design: hardware, firmware, system software, and application software. It is concerned with the central problems facing a supercomputer designer working in today's technologies:

• Physical Design
• The use of commodity silicon to build the system
• High frequency (GHz) signalling on printed circuit boards, through system interconnects, and in copper and fiber optic cables
• Managing the dissipation and removal of heat in the face of very dense system packaging
• Managing electromagnetic radiation and interference caused by high frequency signalling
• Achieving system-level reliability in the face of astronomical component count and regular component failure
• Direct attachment to high-speed communication channels for high-bandwidth streaming of real-time data to and from the machine
• Software and Programming Model
• The derivation of programming models that simultaneously deliver high performance and high programmer productivity
• The gentle degradation of program performance in the face of component failures in the system
• The relative independence of programming model from physical geometry of the machine
• Robust and intuitive mechanisms for diagnosing and debugging the system in software
• The smooth and efficient interoperation of the machine with the prevailing standards for networking and communication

Participants

ILLIAC 6 is a collaboration among a number of organizations and individuals in industry, at the University of Illinois at Urbana-Champaign, and at other Universities. The project is in its early stages and we are actively seeking additional students, participants, and collaborators. Please contact Luddy Harrison if you are interested in working with us on ILLIAC 6.