Building a Teraflop (one million Megaflop) Windows NT Supercomputer
	    and its Implications for Distributed Software

		      Professor Andrew A. Chien
		  Department of Computer Science and
	   National Center for Supercomputing Applications
	      University of Illinois at Urbana-Champaign

The rapid maturation of desktop and high volume technologies in
microprocessors, high speed networks, and operating systems make
commodity building blocks powerful elements for building
supercomputing systems.  In my research group, we have been
researching core system software technologies to synthesize
supercomputers from commodity hardware.  These include lightweight,
low-latency networking software (Fast Messages), coordination
processor and memory scheduling (Dynamic Coscheduling), and efficient
implementations of commodity API's which are combined into High
Performance Virtual Machines (HPVMs).  Delivered HPVM application
performance includes: Message Passing Interface (MPI) with latencies
of 15 microseconds, and 80 megabytes/second, Global Memory (Shmem
Put/Get) and Global Array Interfaces with latencies of 4 and 27
microseconds, and sustained bandwidths of >60 megabytes/second, and
Remote Procedure Call (RPC, Java RMI, DCOM) performance of 47
microseconds (RT) and 35-70 megabytes/second.  HPVM software runs on
both Windows NT and Linux.  Though significant software challenges
remain, these technologies have revolutionary implications for
supercomputing and local-area distributed computing.

To transfer demonstrate these technologies, we have built a series of
clusters (30-processor Pentium Pro, 92-processor Pentium II).  We are
now building a series of high performance Windows NT clusters with
NCSA.  The HPVM/NCSA cluster (April 1998) will include 128 300Mhz
Pentium II processors, and have aggregate capabilities of 38.4
Gigaflops, 32 Gigabytes of DRAM, 10 Gigabytes/second bisection
bandwidth, and 256 Gigabytes of disk storage.  The second HPVM/NCSA
cluster (late 1998) will be 512 Deschutes processors with aggregate
capabilities of >220 Gigaflops (and 4-6x in the other dimensions).
Teraflop systems will follow in 1999.

ANDREW A. CHIEN's research involves networks, network interfaces, and
the interaction of communication and computation in high performance
systems.  His work also involves compilation techniques for high
performance object systems.  He is an Associate Professor of Computer
Science at the University of Illinois with joint appointments with
Electrical and Computer Engineering and as a Senior Research Scientist
in the National Center for Supercomputing Applications.  Andrew
received his undergraduate, master's, and doctoral degrees from the
Massachusetts Institute of Technology and is a recipient of a 1994 
National Science Foundation Young Investigator Award.