Fraunhofer GPI-Space – Parallel Programming Development Software
Domain users are not expert at programming modern high performance computing systems. This statement implies a great many related things: in terms of the energy costs alone, today’s computing systems can cost millions of euros per year. The energy costs in fact, have become a significant share of the total overall cost of a computer. In return, these computers provide incredible possibilities in the form of available computational operations. By the end of this decade, computers will exist that can perform 1018 (one followed by eighteen zeros) calculations per second, that is a quintillion or, a million times a billion operations in a single second.
This huge number describes the theoretical maximal performance, and that is precisely where the problems begin for the domain users. Many fields of science could benefit from simulations that can only be carried out by means of such a powerful capacity, but the computer does not exist. Today’s systems instead consist of thousands of units, each containing thousands of processing cores, linked together via networks. Added to these are special subunits that communicate with the other components over various communication channels.
Each individual processing core must be carefully programmed: for example, the cores can perform more than one operation per cycle, but only if the data to be processed is presented in the proper sequence. A failure to accomplish this means that only 12 percent and, for the newer cores, only 6 percent of the available capacity is used – with no reduction to the cost. The situation with the specialized components is even more complicated and the potential loss in efficiency if not optimally programmed is unacceptably high […]