Printer Friendly Version (spec_article_sharcnet.html)

The Hamilton Spectator
Wednesday, December 12, 2001

 

McMaster getting help untangling the universe

 

by Steve Buist
Science Reporter
The Hamilton Spectator
Dr. Hugh Couchman and Dr. Erik Sorensen

GARY YOKOYAMA, THE HAMILTON SPECTATOR

Dr. Hugh Couchman, left, and Dr. Erik Sorensen are astrophysicists in McMaster University's department of physics and astronomy who will work with the supercomputing network called SHARCNET.

 

Look up, look wayyyyyyyyy up, into the night sky.

To you, it might just look like the moon, some twinkling stars and a lot of inky blackness.

To Dr. Hugh Couchman, however, it represents a lifetime of riddles waiting to be solved.

Couchman is an astrophysicist in McMaster University's department of physics and astronomy.

When he peers deep into the cosmos, his goal is to understand how galaxies formed billions of years ago.

To do so, Couchman sets up three-dimensional mathematical simulations of the universe.

It's nothing that can't be handled by a computer, some complex math equations and the physical laws.

The same physical laws first set out by Isaac Newton more than 300 years ago.

As strange as it might sound, the limiting factor in this marriage for Couchman hasn't been the math or the physics.

It's been the computer.

But that's about to change.

McMaster is joining four other southern Ontario universities and two colleges to create a supercomputing network that will give scientists a much faster and more powerful tool to carry out research.

The same physical laws first set out by Isaac Newton more than 300 years ago.

As strange as it might sound, the limiting factor in this marriage for Couchman hasn't been the math or the physics.

It's been the computer.

But that's about to change.

McMaster is joining four other southern Ontario universities and two colleges to create a supercomputing network that will give scientists a much faster and more powerful tool to carry out research.

It's called SHARCNET -- Shared Hierarchical Academic Research Computer Network -- and it will link clusters of high-performance computers at McMaster, the University of Western Ontario, the University of Guelph, the University of Windsor, Wilfrid Laurier University, Fanshawe College and Sheridan College.

McMaster, Western and Guelph will be the primary centres in the network. All three have recently built supercomputers that rank among the top 400 in the world.

In fact, McMaster had recruited Couchman from Western two years ago in part to launch the university's high-performance computing initiative.

Born in Kent, England, Couchman earned his PhD from prestigious Cambridge University, and spent three years in the astronomy department at the University of Toronto before moving to Western.

There was a time when building supercomputers was a major undertaking. When the world's first specialty supercomputer manufacturer went into business, the company projected that it would sell 10 of its units. Period.

Now, researchers can take advantage of the enormous strides that have been made in computer technology. Just one processor in McMaster's cluster is faster than the first supercomputer that was produced in the '70s.

"If you look at how rapidly processors get faster, realistically, the useful lifetime for the system we have is three or four years," said Couchman.

"Is it just a pile of junk at the end of four years and we've gained nothing? The key point is that we're putting in place a high-performance computing culture."

This first step at McMaster is much like the Field of Dreams philosophy: If you build it, they will come.

Those starting on the ground floor here today can teach the graduate students who will go on to become tomorrow's faculty members.

"It's training people to understand how to use these systems -- what questions you can ask of these systems -- that's important," he added. "It's not the particular flavour of processor or how many megahertz it is that's important.

"The turnover of these things is so rapid that providing the hardware is almost just an operating cost."

Couchman's efforts are already bearing fruit.

A recent addition to the McMaster physics and astronomy department is Dr. Erik Sorensen, who specializes in computational physics and was attracted to McMaster because of the supercomputing possibilities of SHARCNET.

The Danish-born professor has earned degrees from a university in Denmark, the University of California-Santa Cruz and UBC. Most recently, he taught at a university in Toulouse, France.

Sorensen is the first recipient of a SHARCNET research chair.

"The trick in doing high-performance computing is really to figure out the calculations that are possible to do in a reasonable time and then figure out how to do them," said Sorensen.

"There are many, many interesting questions that you could ask but unfortunately there are still many that we can't answer. The trick is to select the right ones that give you information you can learn from."

McMaster's Beowulf cluster, as it's called, is made up of 28 blocks with four computer processors each, for a total power of 112 processors. The beauty of McMaster's new system is that each of the processors is not much different than a high-end Pentium 4-style computer processor that can be bought off a store shelf now.

"The idea behind Beowulf clusters is to go the other way around the problem and profit from the fact that when it comes to computer power versus price, you get so much more if you buy from the consumer market," said Sorensen.

The key has been to figure out a high-speed communication network that can keep up with the processors. To solve the large, complex equations that require a supercomputer, scientists find ways to write them so that the work gets divided up between all the processors. But that presents a new challenge -- some of the calculations are dependent on each other. Each change in one segment of the equation can influence what's happening in another part of the equation.

The problem has to be parcelled out in such a way that the processors are working simultaneously, but each part knows what the others are doing.

McMaster's 28 blocks are all tied together through a switching system that acts like a traffic cop -- this processor needs to talk to that one, and that one over there needs to talk with this one.

"It means you're constantly communicating between processors and that's where the difficulty comes in," said Couchman. "You could spend all your time passing messages and never computing anything.

"For that particular application, you need a very fast network and you need to think very carefully about how these processors talk to each other."

Working at top speed, McMaster's supercomputer can perform 120 billion calculations per second. But the world's fastest supercomputer, at the Lawrence Livermore National Laboratory at the University of California, is still 60 times faster.

The next step will be linking McMaster's 112 processors with the 285 other processors across southern Ontario that will make up SHARCNET.

The real challenge is to then get the fast communication not just between the processors in McMaster's cluster, but also between all the clusters that will be linked across southern Ontario.

Last year, the Ontario government announced that it would provide more than million over five years to develop ORION -- the Ontario Research and Innovation Optical Network.

The goal is to expand the capacity of the high-speed optical fibres that link some of Ontario's universities.

"The problem is, how do we really make this function as one big computer?" said Sorensen. "If you try to do this over the telephone network that most people are familiar with, you know how sluggish it can be.

"If you really want to get going and use high-performance computing, you have to have a very efficient way of connecting.

"The hardware we have here, there's nothing really special about it," he added. "It's hardware that anybody can buy, but the way it's put together is very particular."

Supercomputers are good for exploring two kinds of scientific problems: complex calculations, where the final result might turn out to be one number, and simulations.

With simulations, the problem evolves over time and each step influences the outcome of the next step.

When McMaster's cluster was first installed in the spring, Couchman took it for a test drive to work out the kinks in the system by having one of his postdoctoral students run a simulation problem that looked at the interaction between 250 million elements.

The supercomputer chugged along for two months solving the problem, using all 112 processors. It was equivalent to 10 years' work on a single processor -- but a single processor would never have been able to handle the problem in the first place.

The longest program Couchman has ever run lasted four months.

"If you have to wait much longer than that," he said with a laugh, "the question may have changed before you come to the end."

With SHARCNET, Couchman will now be able to expand the number of elements he can include in his simulations. That should give him more accurate models of how galaxies formed long ago.

And like the universe, the possibilities are endless, which has long been the fascination with astronomy.

So how does an astrophysicist handle the overwhelming vastness of the cosmos?

"I'm afraid the answer is rather prosaic," said Couchman. "You map the universe into some manageable size, you apply equations to that and you don't really think about it.

"The time I really feel that sense of scale is when I give a popular talk, or teaching undergraduate astronomy, when you're trying to describe it.

"OK, here we are on this planet," Couchman continued, "a rather ordinary-looking planet, and as far as we know there's nothing particularly special about it. It's just one of a hundred million stars in the galaxy and that one galaxy is one of billions of billions of galaxies in the universe.

"In that sense, the magnitude of the problem does become a bit difficult to grasp. But you don't really think about that when you press 'Return' on your computer and the program starts running."

You can contact Steve Buist by e-mail at sbuist@hamiltonspectator.com or by telephone at 905-526-3226.

 

 

World's Supercomputer rankings

SITE
Max.
speed
  SITE
Max.
speed
1. Lawrence Livermore National Laboratory, U.S.
7,226
  8. Los Alamos National Laboratory
1,608
2. Pittsburgh Supercomputing Center, U.S.
4,059
  9. Naval Oceanographic Office, U.S.
1,417
3. National Energy Research Scientific Computing Center, U.S.
3,052
  10. Deutscher Wetterdienst, Germany.
1,293
4. Sandia National Labs, U.S.
2,379
  123. Meteorological Service of Canada, Canada.
241
5. Lawrence Livermore National Laboratory, U.S.
2,144
  183. Univ. of Western Ontario, Canada.
185
6. Los Alamos National Laboratory, U.S.
2,096
  280. Univ. of Guelph, Canada.
133
7. University of Tokyo, Japan.
1,709
  315. McMaster University, Canada.
121

 

NOTE: Maximum computer speed is measured in billions of arithmetic operations which can be performed per second.
Source: University of Tennessee/Mannheim University joint project on supercomputer sites.

 

© 2001 The Hamilton Spectator. All rights reserved.
Doc.: 20011212HS508601
This material is copyrighted. All rights reserved. © 2003 CEDROM-Sni