Parallel Architectures research Group – GAP

The Parallel Architectures research Group (GAP) of the Polytechnic University of Valencia has a long experience working on interconnection networks for parallel computers, ranging from large supercomputers, through clusters of personal computers, usually used as servers, to the networks on chip in multicore processors. GAP has also developed research in related subjects, such as, processor microarchitecture and the cache coherence protocols.

The general mission of GAP is to develop applied research of quality, but under realistic hypotheses and considering the market trends, so that enables technology transfer to the industry. This froved effective with the development of fully adaptive routing algorithms in experimental parallel machines (Reliable Router of the Massachusetts Institute of Technology) as well as in commercial machines (Cray T3E of Cray Research Compaq Alpha 21364). More recently, these techniques have been used in the design of IBM BlueGene/L supercomputer (the supercomputer which has headed the Top500 list, the four years, until June of 2008).

Other remarkable hits of the group, measured by their impact in the industry, are the development of very efficient determinist routing algorithms for multistage networks, the development of very compact and versatile routers for networks on chip, and the development of efficient techniques for congestion control. Regarding to the first result, Sun Microsystems has implemented a partitionable version of this algorithm in its Magnum switch, which is the largest switch for InfiniBand at present (3456 ports), and it has been used, among others, in the Ranger supercomputer, that occupied the fourth place of the Top500 list in June of 2008. The second result has been developed in collaboration with AMD; proposing solutions for all the requirements of networks on chip that AMD anticipates for their future processors, so having good perspectives for its practical application. The third result was developed in collaboration with Xyratex, it was patented, and the standardization process of these congestion control techniques began, denominated RECN, like part of the specifications of Advanced Switching Interconnect (ASI).

GAP has been one of the pioneer groups at international level, regarding the research of commercial networks switches (InfiniBand, Quadrics, ASI). In summary, the GAP proposed the first strategies to implement the routing tables in InfiniBand and several routing techniques designed for network load balancing, including some enhancements to use adaptive routing, not originally supported in the standard.

But the main achievement in this field has been the development and standardization of new HyperTransport specifications, in which the number of addressable devices in the network extends, passing from the previous value of 32 to the present one of 4 billion, and defining the new communication protocol and the packet format for this extension, while the compatibility with the previous standard remains. This standarization process has been complex, as supposed to convince companies like AMD, Hewlett-Packard and Sun Microsystems to assume drastic changes in their servers’ architecture, varying from distributed memory architecture to a shared memory, and also how this type of architecture was going to have a commercial demand.    In addition, our group has strong collaboration links with several research groups round of the world (Simula Research Laboratory in Oslo, University of Heidelberg, University of Ferrara, University of Catania, Georgia Institute of Technology, University of Southern California, University of Illinois at Urban, Ohio State University, Pennsylvania State University, Los Alamos National Lab., Princeton University, Keio University); as well as with different companies with cutting-edge technology (IBM, Sun Microsystems, Intel, AMD, Xyratex, Quadrics).