As computational power grows from peta- to exa-scale — a three order of magnitude change that promises a thousand time increase in performance — parallel processing problems become more complex and irregular.
“How to dynamically schedule work for greatest efficiency and how to manage the power consumed in the underlying systems are the key challenges,” explains Wilfred Pinfold, director of extreme scale programs at Intel Labs.
That challenge is one of several being tackled by the Intel-led X-Stack project and ET International Inc. (ETI), a spin-off company with close ties to the University of Delaware’s (UD) Computer Architecture and Parallel Systems Laboratory (CAPSL).
Programming models, languages and related technologies that have sustained HPC application software development over the past decade are becoming antiquated and inadequate for exa-scale era computers, according to Guang Gao, ETI founder and Distinguished Professor of Electrical and Computer Engineering at UD.
This increased complexity requires new thinking and architectures that are portable and sustainable across future generations of computers. Specifically, it is critical that they incorporate energy-efficiencies and resiliencies that allow the technology to transfer as the field continues to advance.
(CAPSL) will develop a self-aware operating system model to reduce energy consumption and save power on these extreme-scale systems. This self-aware operating system will use a novel control model and methodology that employs machine learning to help the system adapt to its environment and effectively “turn off” unnecessary switches as needed to reduce energy consumption.
“If we cannot use these large, complex systems efficiently they will be outside the budget limitations of most institutions. This foundational research is needed to bring these computers within range where we can build them,” says Pinfold.