I-WAY update

by Holly Korab

What began as a glimpse of the future of networking is well on its way to proving that high-speed distributed metacomputing can be a reality. I-WAY, or Information Wide Area Year, the experimental high-performance network scheduled to debut at SC'95, is now connecting more than 17 research institutions across the U.S. I-WAY also is scheduling runs across computing platforms and, in early trial runs, is running parallel codes across different architectures in real time.

Synthesis: I Can Taste What I'm Hearing, Rita Addison, University of Illinois at Chicago, EVL.

"We've come a long way," says Ian Foster, a computer scientist at Argonne National Laboratory. Foster's group is creating standardized software libraries for I-WAY. "This is the first time people have provided a uniform interface to such a heterogeneous group of machines," he says. "It is another step toward the concept of distributed metacomputing."

I-WAY is linking virtual environments, datasets, and computers connected by nine networks of varying bandwidths, protocols, and routing and switching technology. Each node on this network is a supercomputing center with a combined peak computing power of more than a teraflop. Over this experimental network will run approximately 60 different large-scale scientific applications that are part of the SC'95 GII Testbed and HPC Challenge events as well as the SC'95 research exhibits. Each application will use one or several high- performance computers located around the country and will be visualized on the SC'95 conference floor [see access, Summer 1995].

Virtual Molecular Environments on Wide Area Network, Paul Bash, Argonne National Laboratory.

This kind of distributed computing -- in which computers separated by many miles are networked and operated as easily as if they were one single computer -- is thought to be the future of HPCC. It has the obvious advantages in assembling massive amounts of computing potential quickly and inexpensively. In addition, it demonstrates the next wave in telecommunications -- true interoperability between vendor and carrier on a wide scale. Nearly all the communications links involved are OC-3 (155 Mbps) rates or higher, running ATM protocols. Networks utilized include vBNS, AAI, ESnet, ATDnet, CalREN, CASA, ACTS, NREN, and MREN.

Since the first organizational workshop last April, the consortium of I-WAY participants have been standardizing features of this network as well as adopting new technology to support the computing applications that will be coming down its pipeline. Here's where they were in October, six months later. (See http://www.iway.org.)

QMView and GAMESS: New Insight into Molecular Structure and Reactivity through High-Performance Scalable Processing, 3D Visualization, and High-Speed Networking,Kim Baldridge, San Diego Supercomputer Center.

I-POPs

To help standardize the I-WAY, the 17 key sites -- those that have volunteered their resources for the I-WAY and SC'95 -- are installing point-of-presence (I-POP) computers as their gateway to the I-WAY. The I-POP machines are configured uniformly and possess a standard software operating environment to overcome issues of heterogeneity, scale, performance, and security. The I-POPs schedule applications and serve as proxies to establish jobs. For instance, researchers who will be using NCSA's computers will not log directly into the system; rather, they will log into the center's I-POP, which in turn will initiate the run. I-WAY participants are assigned port numbers and tokens that are similar to passwords but are only good for specific port numbers and time frames. If connections need to be established with resources at other centers -- say if the gravitational field being calculated on NCSA's HP/CONVEX Exemplar also will be running on Cornell's IBM SP2 -- the I-POPs will perform authentication at the second site as well as "trusted services." This network of I-POPs reduces dangers inherent in distributing passwords. "What they have done is made sure that only an authorized individual can start up the process," says Ken Rowe, NCSA computer security coordinator.

Currently there is no reservation system. That will be added to ensure bandwidth requirements between sites by December, when an estimated 60 applications will be packed into the four days of the conference. "Right now we just want people to get on the machines and test their applications," says Foster. When researchers log onto an I-POP (using pre-established commands), they are told what other applications are queued and when time will be available. They also learn the status of the testbeds.

Real-Time and Near-Real-Time Space Weather Forecasting Charles Goodrich, University of Maryland and Pittsburgh Supercomputing Center.

Testbeds

By October, researchers could choose from among four testbeds grouped according to their computing architectures. They were the Alpha, Argonne's IBM SPs and Cornell Theory Center's SPs; Beta, Caltech's Intel Paragon and San Diego Supercomputer Center's Paragon; Gamma, Cornell's SGI POWER CHALLENGE and NCSA's POWER CHALLENGE ARRAY; and Theta, Argonne's SPs and the CRAY T3D at Lawrence Livermore National Laboratory. The Alpha testbed has been up since July, when researchers first demonstrated the capability of the I-POPs to spin off a job. The composition of these testbeds may change before SC'95; defining these subsets simplifies scheduling for trial runs.

Community Climate Model 2: Climate Simulation Lab, Don Middleton, National Center for Atmospheric Research.

Proof of concept

The question of greatest concern to the scientists was partially answered in July when Joan Masso, Ed Seidel, Rob Gjertsen, and Mark Nardulli from NCSA's Relativity Group showed that distributed computing can work. To test Message Passing Interface (MPI) -- the library of images that enables data to be ported across architectures -- they ran 3D computational code for solving the general set of Einstein equations for a gravitational field on a single parallel machine that treated each of its 512 nodes as a computer. This required breaking up the computation 512 ways. In July they ran one calculation on three machines at once: IBM SP2, IBM SP1.5, and SGI POWER CHALLENGE ARRAY systems in truly distributed fashion. The process was inefficient because the network lines they used were undedicated, but it did provide proof of concept.

Cells and Smaller: Exploring the Machinery of Life, Richard E. Gillilan, Cornell Theory Center.

"We know it can work," says Seidel. "But it won't be until the actual day of the conference -- with high-level government dignitaries standing there -- that you know it will work. Then you will see all five experimental machines running across an experimental network. Just think about how difficult it sometimes is to get one computer working right. Imagine five."

Cellular Semiotics: Molecular Recognition on Biological Membranes, Marcus Wagner, NCSA

Holly Korab is a science writer in the NCSA Publications Group.

Return to the Table of Contents.

NCSA Home Page

access / Fall 1995 / Email comments to NCSA Publications Group: pubs@ncsa.uiuc.edu