FOR RELEASE: Wednesday March 22nd, 9:48am.
Dave Turner
(515-294-8872,
turner@ameslab.gov)
Ames Laboratory - ISU
Ames, IA, 50011
Popular Version of the Paper
K6.004
Wednesday March 22nd, 9:48am.
APS March 2000 Meeting, Minneapolis
Today's largest supercomputers, carrying exotic names such as the Cray T3E, SGI Origin2000, and IBM SP, can cost millions of dollars and often fill an entire room. They may have thousands of individual processors connected by a custom communication network that allows them to work together on a single task. It takes thousands of Amps to power them, and tons of air conditioning to keep them cool. Most are esthetically beautiful, with sleek designs and enough blinking lights to make them look right at home in any science fiction movie.
There is a broad range of science performed on these computational marvels. The very largest of these supercomputers are set aside primarily for simulations that insure the readiness of the nations nuclear weapons arsenal. These simulations provide the US government with the confidence necessary to continue with the Nuclear Test Ban Treaty.
The entire combustion process inside an engine is being modeled to produce more efficient fuel mixtures and piston designs. This is a new type of simulation that requires multiple time and length scales to be performed all within a single simulation. At the lowest level, the fuel-air mixture is ignited causing chemical reactions to occur. This in turn heats the gases within the entire chamber causing them to expand and move the piston. The chemical reactions must be simulated on very fine length and time scales, but must also be tied into the larger scale simulation of the entire piston in an accurate and efficient way.
Specialized supercomputers such as IBM's Blue Gene will be used to understand some of the basic processes of life. Proteins control all chemical processes at the cellular level. Understanding how proteins fold from single long strands into complex 3D structures is helping scientists determine their final shape, and therefore their function. Some diseases occur from small changes in the folding process, so understanding this process can lead to an understanding of the disease, and eventually to a treatment or cure.
There are many projects on a similar scale called 'grand challenge' projects. The Condensed Matter Theory group at Ames Laboratory, a Department of Energy laboratory run by Iowa State University, has been involved in several projects that have required a large effort to develop efficient computer algorithms, and require large amounts of computer time to undertake. This materials science research is focused on trying to understand the basic properties of materials, which can lead to designing materials with better properties. Simulations may involve looking at the dynamic properties of millions of atoms or doing a more accurate calculation on only a few hundred. Projects may involve investigating the growth of thin films on silicon wafers, or studying microscopic marvels such as small carbon Buckyballs. The photonic band gap project studies the propagation of electromagnetic waves as they pass through a medium. This is being used to design optical waveguides that direct light between optical components that may be part of an optical computer in the future.
While the computational power of these supercomputers is impressive, their cost is great due primarily to the custom networking that connects the processing nodes. The idea of building a 'supercomputer' by connecting many workstations or PCs together in a cluster with high speed networking is clearly attractive. After all, todays workstations use the same processors that go into many of these supercomputers. The main shortcoming of these 'cluster computers' is that the interprocessor communication rate can be 5 to 10 times slower than in a traditional supercomputer. This makes it much more difficult to spread a program across the processing nodes efficiently, limiting the type of applications that can be run on workstation clusters.
If we can bring the performance of the communication network closer to the level found in traditional supercomputers, we can build cluster computers that rival supercomputers but are a fraction of the cost. The network cards and switches have improved to a level that they can handle data as fast as most workstations can dump it out. The internal memory bus is the main bottleneck, but even this is improving rapidly as computer designers realize that faster busses are needed to feed the fast processors.
A great deal of the raw performance is lost in the software layers. Each parallel program uses message passing calls to exchange data between the individual processing nodes. The message passing library is the software that handles all the network traffic on a given system. There are many such libraries, most of which adhere to a standard called the Message Passing Initiative (MPI). A program that uses one of these MPI libraries will run on a wide variety of computer architectures without alteration. However, with most MPI libraries this portability comes at the price of performance.
The author has developed a streamlined MPI implementation called MPI_Lite that provides the maximum performance while maintaining portability. This library speeds up the communication rate by 30% to 100% on the machines supported. This same library supports channel bonding, a technique where each computer can communicate over more than one network interface simultaneously. For a PC cluster connected by Fast Ethernet, this can double the network performance without increasing the cost much.
While more improvements are needed in the hardware capabilities, many of the software impediments are quickly disappearing. Cluster computers are rapidly becoming more mature systems capable of running a wide variety of scientific codes. A few more years of hard work and improvements to the hardware is all that is needed to realize the dream of building a true supercomputer out of commodity components.