Supercomputers belong to a class of highly specialised hardware infrastructures, where high number of machines are typically pre-organised and smart-linked together with specialised high-speed low-latency interconnects, so as to allow new forms of concurrent processing cooperations to be orchestrated. Having any such supercomputing infrastructure is not enough, it is important to also use system tools capable to harness the most of the available CPU-powers
Supercomputers first began to appear in the 1960's.
These early supercomputers had only a single, high-speed processor. Control Data Corporation's CDC-6600, designed by Seymour Cray, was about ten times faster than all other computers of its day, and was dubbed a supercomputer -- the first appearance of the term.
Later, as processing speed, cooling ability, and physical size hit limits, Cray pioneered the method of linking multiple processors together in order to get more speed out of the same machine. This is the same method used in today's supercomputers, which can range in size from thousands of processing cores to hundreds of thousands of processing cores.
* Seymour CRAY ( yes, the supercomputer guy )
* said:
* --------------------------------------------------------------------
* A supercomputer turns compute-bound problems into I/O bound problems
* --------------------------------------------------------------------
* and:
* --------------------------------------------------------------------
* It is not hard to build a fast processor or a fast memory,
* but the challenge is to build a fast system.
* --------------------------------------------------------------------
Interconnect latency is an additional [TIME]-domain penalty, each process has to pay for using a supercomputer's remote resource under a distributed computation-graph schedule.
Minimising interconnect's latency-costs is thus one natural direction, using a smarter, overhead-aware computation-graph design is the other direction to achieve the indeed I/O-bounds' bleeding edge of the ultimate performance from any supercomputing system's infrastructure.
