Supercomputing Is Here!

For some colleges and universities, computing is now soaring into the stratosphere.

Mike Hickey

Mike Hickey at Embry-Riddle Aeronautical University
is using a brand-new supercomputer to learn
more about acoustic-gravity waves in the upper
portions of Earth’s atmosphere— waves which
ultimately impact flying conditions.

When most academic technologists tackle computing, their time is occupied by laptops and servers—relatively small-scale stuff. A couple of blades here, a couple of blades there. Generally, even for network managers, the processing power rarely stacks up to anything awe-inspiring. Sometimes, however, computing can be bigger and broader than many of us can imagine, requiring more juice than some small nations use in a year. Then, of course, we find ourselves in the 21st-century realm of highperformance computing. High-performance computing efforts at four schools—Indiana University, the University of Florida, the University of Utah, and Embry-Riddle Aeronautical University (FL)—demonstrate that the latest and greatest in supercomputing on the academic level far exceeds the computing power that most of us can conceive. As computing power continues to grow, however, these tales undoubtedly are only the beginning.

Forecasting Improvements

In the wake of the devastation caused by Hurricane Katrina, a number of today’s most elaborate high-performance computing endeavors revolve around finding better ways to predict everything from everyday showers and snowstorms to devastating hurricanes and tornad'es. At Indiana University, researchers in the School of Informatics are using their high-speed computing and network infrastructure to help meteorologists make more timely and accurate forecasts of dangerous weather conditions. The project, which kicked off with an $11 million grant from the National Science Foundation in 2003, is dubbed Linked Environments for Atmospheric Discovery (LEAD).

The LEAD system runs on a series of remote, distributed supercomputers—a method known as grid computing. Co-principal Investigator Dennis Gannon, who also serves as a professor of Computer Science at the university, says the project is designed to build a “faster than real time” system that could save lives and help governments better prepare themselves for looming natural disasters. Today, the project is in its infancy—as yet in the planning and testing stages. Ultimately, however, Gannon sees the model outpacing the current strategy for weather prediction—a system that, despite constant improvements, still runs largely on simulations.

“Our goal is nothing short of building an adaptive, ondemand computer and network infrastructure that responds to complex weather-driven events,” he says. “We hope to use this technology to make sure storms never cripple us again.”

The LEAD system pools and analyzes data received from other sources such as satellites, visual reports from commercial pilots, and NEXRAD, a network of 130 national radars that detect and process changing weather conditions. Down the road, an armada of newer and smaller ground sensors dispatched to detect humidity, wind, and lightning strikes will be part of the network, too. As weather information comes in, it is interpreted by special software agents that are monitoring the data for certain dangerous patterns. Once these patterns are identified, the agents will dispatch the data to a variety of high-performance computers across private networks for real-time processing and evaluation.

In most cases, these collection devices will send weather data out for processing on computers in the IU network. Sometimes, however, spurred by an additional $2 million grant from the National Science Foundation (NSF), the software agents will dispatch data to computers on a broader distributed computing network known as TeraGrid. The grid is a national network that allows scientists across the nation to share data and collaborate. Under this system, huge computers in San Diego, Indiana, and Pittsburgh are linked via a 20GB connection rate, to facilitate cooperation. The result: several thousand processors at a participating school’s fingertips, on demand.

“My files may come from San Diego and my computing facility may be in Pittsburgh, but the network is such that the facility in Pittsburgh d'esn’t care where the data’s from,” explains Gannon. “As you can see, this kind of leverage opens up a host of new doors in terms of what kind of weather data we can process with supercomputers, and how we can do it.”

The University of Florida’s new cluster boasts tightly coupled nodes that act like a single computer; no more disparate calculations attempting to handle huge computations.
Clusters: Loose or Tightly Coupled?

The science of supercomputer processing is far from easy. Generally, the “engines” of supercomputers are gaggles of processing power called “nodes.” Each node is comprised of a series of processors, and each processor differs from the others depending on how sophisticated the supercomputer is. By and large, these nodes usually boast two to four central processing units (CPUs) with up to 4GB of RAM. To put this into perspective, the best nodes basically are the same as four really expensive personal computers. And most supercomputers have at least 100 of these nodes—the equivalent of 400 of the fastest and most efficient computers money can buy.

SupercomputingStill, not all supercomputers are created equal. Technologists at the High Performance Center (HPC) at the University of Florida recently unveiled a brand-new cluster: a 200-node supercomputer that’s bigger than anything the school has had before. With the speed of this machine, HPC Director Erik Deumens says UF researchers will embark on new projects to investigate the properties of molecular dynamics, the ins and outs of aerodynamic engineering, and climate modeling projects of their own. Deumens calls the bulk of these projects “multi-scale”—an approach that takes into account mathematical

“You try to describe a certain piece of a problem with a particular methodology, but you know that in some part, something more interesting is happening, so you use the magnifying glass of advanced calculation,” he says, pointing to one researcher who is studying the molecular interaction of heated silicon when engineers etch microchips. “The only way to make sure you’re not overwhelmed with numbers is to look at problems with enough computing power to answer multiple questions at once.”

Deumens notes that with a 200-node supercomputer, connections between nodes are critical. To make sure the machine functions properly, HPC turned to Cisco Systems for all of the networking connections between nodes. According to Marc Hoit, interim associate provost for Information Technology, the vendor also is helping UF connect all of its clusters on campus so HPC can perform more gridbased computations. All told, Hoit estimates that soon, more than 3,000 CPUs will be part of this grid. UF also will contribute to the Open Science Grid, an international infrastructure in the vein of TeraGrid, though considerably larger.

One key differentiator between the expanded grid and UF’s new cluster is the way in which the nodes are coupled. In the UF cluster, the nodes are more tightly coupled, meaning that they act more like one single computer, and can be harnessed to perform large-scale mathematical computations quickly, as one unit. In the grid, however, the nodes are loosely coupled, meaning they all have other computing responsibilities, and likely are not ready to handle huge calculations at one time. The “loosely coupled” strategy is perfect for small equations with small sets of data, such as genetic calculations. For weather and aerodynamic simulations, however, Hoit notes the tightly coupled approach is a must.

“Many people often make simple statements about CPUs, and say they can apply all the power to solving one large problem,” he says. “But a tightly coupled approach can help achieve efficiency for large problems, too.”

WHAT IS MYRINET?

Myrinet is a networking system designed by Myricom; it has fewer protocols than Ethernet, and so is faster and more efficient. Physically, Myrinet consists of two fiber optic cables, upstream and downstream, connected to the host computers with a single connector. Machines are connected via low-overhead routers and switches, as opposed to connecting one machine directly to another. Myrinet includes a number of fault-tolerance features, mostly backed by the switches. These include flow control, error control, and “heartbeat” monitoring on every link. The first generation provided 512 Megabit data rates in both directions, and later versions supported 1.28 and 2 Gigabits. Newest “Fourth-generation Myrinet” supports a 10 Gigabit data rate, and is interoperable with 10 Gigabit Ethernet. These products started shipping in September 2005.

Myrinet’s throughput is close to the theoretical maximum of the physical layer. On the latest 2.0 Gigabit links, Myrinet often runs at 1.98 Gigabits of sustained throughput—considerably better than what Ethernet offers, which varies from 0.6 to 1.9 Gigabits, depending on load. However, for supercomputing, the low latency of Myrinet is even more important than its throughput performance, since, according to Amdahl’s Law, a high-performance parallel system tends to be bottlenecked by its slowest sequential process, which is often the latency of transmission of messages across the network in all but the most embarrassingly parallel supercomputer workloads.

And Now for the Metacluster…

Running a giant cluster like the one at UF is remarkable, but imagine operating something five times that big. Such is life for Julio Facelli, director of the Center for High-Performance Computing (CHPC) at the University of Utah. The Center is responsible for providing high-end computer services to advanced programs in computational sciences and simulations. Recently, via a grant from the National Institutes of Health (NIH), CHPC has purchased a metacluster to tackle the new generation of bioinformatics applications, which comprise the nitty-gritty study of genetic code and similarly complicated equations. Facelli says the machine is one of the largest of its kind in the academic world.

The new metacluster boasts more than 1,000 64-bit processors in dense blades from Angstrom Microsystems. The metacluster has been configured into five subsystems, including a parallel cluster with 256 dual nodes, a “cycle farm” cluster with 184 dual nodes, a data-mining cluster with 48 dual nodes, a long-term file system with 15 terabytes of storage, and a visualization cluster driven by a 10-node cluster. Most of these clusters are connected by Gigabit Ethernet. The parallel cluster runs on Myrinet, a networking system designed by Myricom that has fewer protocols than Ethernet, and therefore is faster and more efficient. (See box above, “What is Myrinet?”)

“When we started this program eight or nine years ago, it was possible only to do simulations in one dimension,” says Facelli. “Today, we are computing in three dimensions and running calculations that we never dreamed of being able to run.”

In addition to these resources, the metacluster boasts a “condominium”- style sub-cluster in which additional capacity can be added for specific research projects. With this feature, Facelli says CHPC uses highly advanced scheduling techniques to provide seamless access to heterogeneous computer resources necessary for an integrated approach to scientific research in the areas of fire and meteorology simulations, spectrometry, engineering, and more. Additional specialized servers are available for specific applications such as large-scale statistics, molecular modeling, and searches in GenBank, a database of genetic data. CHPC is developing several cluster test beds to implement grid computing, too.

Down the road, CHPC plans to add two dual nodes to its “condominium”- style cluster for a proposed study of patient adherence to poison control referral recommendations. As Facelli explains it, the school seeks to use machine learning methods for feature selection and predictive modeling—an enterprising approach, considering that prior to this, no researcher or research institution had ever implemented highperformance computing to accomplish such a challenge. CHPC will support these nodes for the duration of the project and will collaborate with the project investigators in the computational aspects of the research. Afterward, Facelli says, the nodes will be subsumed back into the system.

“You can never have too many nodes in your metacluster,” he quips. “We’re excited about the possibilities of what these will bring.”

WHAT IS BEOWULF?

Aside from being a classic work of early literature, Beowulf is a type of computing cluster. The label is a design for high-performance parallel computing clusters on inexpensive personal computer hardware. Originally developed by Donald Becker at NASA, Beowulf systems are now deployed worldwide, chiefly in support of scientific computing. There is no particular piece of software that defines a cluster as a Beowulf. Commonly used parallel processing libraries include Message Passing Interface (MPI) and Parallel Virtual Machine (PVM), a software tool developed by the University of Tennessee, Emory University (GA), and The Oak Ridge National Laboratory. Both of these permit the programmer to divide a task among a group of networked computers, and recollect the results of processing.

Flying High

Another school that has researchers excited about the future is Embry-Riddle Aeronautical University (FL). There, in the school’s four-year-old Computational Atmospheric Dynamics (CAD) laboratory, Mike Hickey, associate dean of the College of Arts and Sciences, is using a brand-new supercomputer to learn more about acoustic-gravity waves in the upper portions of Earth’s atmosphere. Driving Hickey’s research is a 131-node, 262-processor Beowulf cluster (see box above), which runs simulations of waves propagating through the atmosphere. These waves ultimately impact flying conditions, which is precisely why the research is of such value to a school like Embry-Riddle.

Such simulations used to take three or four days to run; with the power of the new machine, however, Hickey can run them in a matter of hours. Elsewhere at the school, other researchers are turning to Beowulf to speed up projects of their own. Hickey points to a number of plasma physicists who are calling upon the computer to simulate plasma flow and interactions, and engineers who are running simulations of the flow of gasses through turbine engines. One professor even uses the system to analyze information from a database of all the commercial airline flights in the US for the last decade; from this information, the professor is trying to predict flight delays down the road.

“Especially at an engineering school like ours, there’s a lot of numerically intensive simulation work on campus,” says Hickey, who explains that in general, Beowulf clusters are groups of similar computers running a Unix-like operating system such as GNU/Linux or BSD. “The best way around [the demand for so much simultaneous simulation work on one campus] was to try and get a computer that serves everybody’s needs.”

Still, the emergence of Embry-Riddle’s supercomputer has not been without hiccups. The first challenge revolved around program code: Researchers can’t just take the code that runs on a single processor, move it over to the new machine, and run it; instead, code for the supercomputer needs to be heavily modified in order to take advantage of the multiple processing capabilities. Consequently, last summer, Embry-Riddle ran a workshop to educate some professors and graduate students about how to manipulate data for the new machine. Another subject of the workshop: Message Passing Interface (MPI), an architecture that must be learned by a user in order to understand when the high-performance computer is ready to receive new material.

The University of Utah has purchased a metacluster to tackle a new generation of bioinformatics applications; the machine is one of the largest of its kind anywhere in the academic world.

Embry-Riddle officials also have tackled challenges of a more logistical nature. CIO Cindy Bixler says that when Hickey came to her and requested the supercomputer, the school’s facilities department didn’t understand that investing in a machine of that magnitude would require the school to rethink its server room completely. Computer clusters run hot, so the school had to invest in additional air-conditioning units. Bixler then brought in an engineer to study the server room’s air flow and figure out where to put the new cooling device. Finally, of course, was the issue of electricity—with the new machine, Embry-Riddle’s energy bills went through the roof.

“You can’t just flip the switch on a high-performance computer and expect everything else to work itself out,” Bixler says. “This kind of effort takes considerable planning, and in order to avoid surprises, schools need to be ready before they buy in.”

comments powered by Disqus