PORTLAND, Ore. Exascale data centers would harness farms of petaflop-caliber computers to achieve 1,000-fold increases over the world's fastest computers.
Hewlett-Packard Laboratories wants to lay the groundwork for exascale computing (an exaflop is equal to 1,000 petaflops) through a three-year grant to Georgia Institute of Technology (Atlanta) to develop Xen-based virtualization management software.
"Our work is designed to virtualize heterogeneous multi-core platforms that include general purpose cores, special purpose nodes like graphics accelerators and the management of tasks," said Karsten Schwan, a professor at the Georgia Institute of Technology and director of its Experimental Research in Computer Systems. "We are extending the open source virtualization software, called Xen, for exascale computing."
The open-source Xen virtualization system adds a layer of software layer called a hypervisor (about 150,000 lines of code) between a server's hardware and its operating system. It provides an abstraction layer that permits each physical server to run any number of virtual servers. By decoupling the operating system and its applications from a server, the hypervisor can manage resources and tasks more freely, relocating running programs even during hardware failures.
"If we can achieve our goal of virtualizing exascale multi-core computers, then many large scale applications can be enhanced, such as weather simulations, with much finer resolutions," said Schwan.
Besides climate modeling, exascale computing could also aid biological simulations, drug discovery, national defense, energy assurance and advanced materials development.
The three-year program will culminate with a software suite that manages a heterogenous mix of different core types, from general purpose computers to graphics accelerators to database systems. Schwan's team plans to virtualize a prototype system composed of several different types of cores to prove that their extension of the Xen management software can handle the job. It will then be up to HP Labs to test the prototype software on exascale caliber data centers.
"Exascale computing will only be achieved by virtualizing general purpose multicore processors with special purpose chips like graphics accelerators. That is the real challenge," said Schwan.
A second grant for exascale computing was awarded by HP Labs to the University of Michigan for its Gadara program aimed at discrete control of deadlock avoidance in concurrent software.