Reconfigurable hybrid interconnection for static and dynamic scientific applications
Shoaib Kamil, Ali Pınar, Dan Gunter, Michael J. Lijewski, Leonid Oliker, John M. Shalf · 2007
As we enter the era of peta-scale computing, system architects must plan for machines composed of tens or even hundreds of thousands of processors. Although fully connected networks such as fat-tree configurations currently dominate HPC interconnect designs, such approaches are inadequate for ultra-scale concurrencies due to the superlinear growth of component costs. Traditional low-degree interconnect topologies, such as 3D tori, have reemerged as a competitive solution due to the linear scaling of system components relative to the node count; however, such networks are poorly suited for the requirements of many scientific applications at extreme concurrencies. To address these limitations, we propose HFAST, a hybrid switch architecture that uses circuit switches to dynamically reconfigure lower-degree interconnects to suit the topological requirements of a given scientific application. This work presents several new research contributions. We develop an optimization strategy for HFAST mappings and demonstrate that efficiency gains can be attained across a broad range of static numerical computations. Additionally, we conduct an extensive analysis of the communication characteristics of a dynamically adapting mesh calculation and show that the HFAST approach can achieve significant advantages, even when compared with traditional fat-tree configurations. Overall results point to the promising potential of utilizing hybrid reconfigurable networks to interconnect future peta-scale architectures, for both static and dynamically adapting applications.