"Fusion Cluster: Reconfigurable, High Performance and Scalable Cluster Architecture"

N Venkateswaran, T P Ramnath Sai Sagar, G Shyamsundar, E Vinoth Krishnan, and K Viswanath
WAran Research FoundaTion (WARFT)

Abstract

When it comes to aerospace applications power aware computing [1] [2] plays an important role. For on-board supercomputers, both power and performance are of equal importance due to mission criticality of applications and hence no trade-offs are acceptable. Many approaches have provided methods of increasing performance at node level. However to increase performance at the cluster level, scalability becomes a crucial factor with respect to performance and power. In the DSM era, development of hardwired clusters proves to be an impediment while designing multi-million node clusters. Although wireless interconnects have insufficient bandwidths and latencies, they also have very high fan-in and unity diameter which aid in the achievement of low power. To tackle these problems we propose a hybrid network of tightly wired ’islands’ (encompassing a group of computation nodes and I/O nodes) and wireless links across these ’islands’. This hybrid configuration is termed as the “fusion cluster”. The wireless aspect incorporates low power data transfer as well as ability to reconfigure the topology of cluster network. Reconfigurability of the topology will facilitate mapping of different applications having varying communication requirements; in particular with respect to aerospace applications that include image processing, deep space astrophysics based modeling and discovery of new space medicines[4] (through computationally intense brain modeling) for correcting brain abnormalities which arise due weightlessness. Such an element of re-configurability will greatly alleviate the cost prohibitive nature of application specific supercomputers and cater to a spectrum of applications while aiding in low on-board power consumption. By balancing the computational and I/O node count, its topology within a wired island, against the number of such wired islands and the wireless communications across them, the Performance to Power ratio [3] scalability can be achieved. 

References: 

  1. Arrvindh Shriraman. Pascom - Power Aware Supercomputing- a supercomputer power model. Thesis submitted to WAran Research FoundaTion (WARFT), Aug 2004. 

  2. Ram Rajamony, Pat Bohrer, Mootaz Elnozahy, Mike Kistler, Charles Lefurgy, and Chandler McDowell. The case for power management in web servers. Power aware computing, Graybill and Melhem eds, Kluwer academic publications, 2002. 

  3. Prem Kumar Ramesh - PetaOPS Years Effect: Temperature Balanced Mapping and Prediction of Application Dependent Device and Interconnect Aging, Thesis Submitted to WAran Research FoundaTion (WARFT), 2005.  

  4. V S; FOMINA G A ATKOV, O I U; BEDNENKO. Ultrasound techniques in space medicine. Aviation, Space, and Environmental Medicine. Vol. 58, pp. A69-A73, Sept.

2006 MAPLD International Conference Home Page