Data Communications in an HPC Hybrid Cluster and Performance Evaluation
Ovidiu Gherman, Ioan Ungurean, Gheorghe Pentiuc, Oana Mihaela Vultur · 2010
The need for powerful computing systems is more present than ever in the today scientific (and not only) environment. Cluster systems built with general purpose processors (a trend that caught in the previous years) are superseded today by platforms built around specialized multicore processors, capable of processing massive amounts of arithmetic operations, usually from the field of graphic accelerators (GPU). These units are specifically created to manage intensive operations so they are the logical choice for this purpose. IBM took this step by creating a platform that uses a multicore accelerator CPU (Cell BE) for the HPC operations, managed by nodes with a more traditional architecture, with general purpose CPUs (AMD Opteron). This hybrid approach toward HPC is successful, even if the application deployment and software development can pose a certain degree of difficulty.