Cluster Computing
In order to improve computing performance, there are two techniques. One is just to improve the performance of a single processor. Another is to prepare multiple processors and divide the program to shorten the execution time according to the number of divisions. Especially in the latter one, an organization in which general-purpose processor units such as so-called personal computers are connected via a network is called a cluster computer environment. The conventional supercomputers developed dedicated components such as processors and memory. But with the recent widespread use of personal computers, the cluster computer has become possible to inexpensively configure the parallel computing environment that connect them with network connections. Many of the high-performance supercomputers these days have such a configuration. A single processor these days has become very fast. On the other hand, the overhead of connecting them to migrate data is often a problem. We are focusing on the problem and developing the next-generation networks.
GPU Computing
Processor performance improvements have been improved by Moore’s Law, a market principle that “the degree of integration of LSIs will quadruple in three years.” However, as the miniaturization of integration technology progressed, and the processor architecture became more complicated, the law began to break down. At the same time, manycore processors have emerged that integrate thousands or tens of thousands of small processors instead of limiting them to some of the features of traditional processors. Manycore processors were originally a processor (GPU) technology used for graphics processing, and have recently become a high-speed computing platform used for scientific, technological calculations and AI. Focusing on high-performance computing using GPU, we are conducting research projects aiming at large-scale parallelization.