Advanced Computing Architectures
Many in Indiana have long understood that the supercomputers of tomorrow must be built on more than just ever-faster commodity microprocessors. New approaches to processing and the linkages of multiple computers together to create advanced computing grids require constant innovation and effort. Simply taking advantage of increasing speeds of commodity microprocessors will not solve the challenges facing us today.
Purdue and Indiana Universities
Indiana and Purdue have long been leaders in the nexus between networking, supercomputing, and grid computing. Purdue and Indiana Universities have now been asked to join the TeraGrid, a nationwide effort to build a national grid of advanced supercomputers, massive data storage systems, and advanced instruments, all joined together by high speed networks to create a nationwide system to perform the nation’s most advanced research – termed “cyberinfrastructure.”
The joint IU-Purdue effort will create a system within the state, called the IP-grid, which will join together the prodigious supercomputing, massive data storage, advanced instrument, and data resources at Purdue University and Indiana University. IU and Purdue already operate an advanced optical-fiber network, called I-Light, which links IU Bloomington, Purdue's West Lafayette campus, and Indiana University – Purdue University Indianapolis. The universities will provide both raw computing power and huge amounts of research data in areas ranging from global weather statistics to retail sales, satellite data to chemical catalysts. Particularly important additions to the TeraGrid include the following:
- Sophisticated visual displays and specialized software for turning
data into three-dimensional images
- A large set of data containing thousands of years of global weather
information
- An entire year's worth of sales-transaction data from one of the
world's largest retailers, information that is useful for economists
and social scientists
- The Purdue Terrestrial Observatory, which will provide real-time
satellite data on quickly evolving changes in the environment during
major storms, earthquakes and even terrorist attacks. The information
could be critical for the first emergency personnel responding to disasters
- A tremendous array of public life sciences data made available through
Indiana University’s Centralized Life Science Data service
- A "crisis grid," in which researchers use simulation software
to predict the likely outcomes of natural and manmade catastrophes,
such as a bioterrorism attack, and develop the most effective plans
to contain such emergencies
- 1 trillion cycles, or teraflops, of computing power on a routine
basis and up to 6.26 teraflops for especially demanding and important
jobs
- Access to more than 400 terabytes, or trillion bytes, of storage capacity
University of Notre Dame
Researchers at the University of Notre Dame have become national
leaders in the development of Processor in Memory (PIM) technology,
an approach to massive computing systems that may revolutionize
supercomputing.
With such technologies, the “memory wall” that represents the single most significant obstacle to using modern technology efficiently is overcome by integrating the processing logic onto the same chips as the memory. Ultimately, this makes for single component systems, consisting of nothing but memory. In such a system, computation is performed in whichever memory chip happens to contain the pertinent data. Older techniques for accessing memory are giving way to new ones that “migrate the computation,” thereby changing the latencies involved with two-way accesses (that is, sending a request and waiting for a response) to very efficient one-way operations. Relentless multithreading throughout such systems allows for levels of concurrency that are orders of magnitude higher than those offered by conventional parallel supercomputers.
For more than a decade, architectures developed at Notre Dame using PIM technology have focused on high-end applications. They have covered the gamut from embedded supercomputers for deep-space probes, through “memory cards” for conventional workstations that have more parallel computing power than the microprocessor access them, to HTMT—a PetaFLOPS-level supercomputer design utilizing a spectrum of advanced technologies.
Most recently, Notre Dame has teamed with Cray Inc. to develop a new generation of trans-PetaFLOPS-level commercial systems. In July 2003 after a successful Phase 1 effort, Cray and its university research partners—Notre Dame, Stanford, and Cal Tech—were awarded $50 million to participate in the research and development phase of the High-Productivity Computing Systems (HPCS) program funded by Defense Advance Research Projects Agency. The HPCS program was formed to sponsor development of the next generation of high-productivity computing systems, systems that are more broadly applicable, much easier to program, and more resistant to failure than currently-available high-performance computing systems. The Cascade project, as this new three-year effort is called, will utilize a spectrum of new architectural techniques. Notre Dame’s PIM architectures a central facet in revolutionizing supercomputing.