Why the top of the pyramid resources are needed ?

In competitive science and research only one thing counts: the time to solution of any particular problem. Therefore sufficient resources for qualified research should be readily available. It also means that High Performance Computing (HPC) cannot be considered without the broader context of all resources for Research Computing in Europe. Research Computing requires high end facilities to fit or match with the European and national research profiles and with the nature of the very application codes that are being used. This applies to the dimensions of each resource and its architecture. The European and national research profiles are determined by general or social interest, emphases placed by research councils and the sheer number of researchers in a particular field. So where does HPC distinguish itself from any other type of research computing? The distinction is defined by the means problems can be solved that are intrinsically latency bound.

Supercomputers, as they were called at the time of single and multiprocessing, have always distinguished themselves by their ability to solve latency bound problems. The first latency hurdle is the time to access memory from the CPU. This still is a crucial factor which has not been overcome yet. Much, however, has been done to hide this latency by the introduction of a layer of memory caches, combined with compiler intelligence. It is expected that new technological advances in chip architecture will further reduce the latency problem on the processor level.

Present day computing facilities, which all belong to a category of massively parallel computers (whether the processors are scalar or vector type and 32- or 64 bits wide), still face a formidable latency problem, namely in the network that connects the processors. Dedicated internal networks are still factors faster than general purpose networks, but still are hardly capable of matching the ever increasing clock frequencies of the processors. And even if they ever would, the latency still would be critically dependent on the distance that the signals will have to travel: no more than 30 cm per nanosecond, due to the absolute limitation by the speed of light in vacuum.

Tags : , , , , , , , , ,

If you enjoyed this post, please consider to leave a comment or subscribe to the feed and get future articles delivered to your feed reader.

Leave Comment