 |
Cellular architecture builds next generation supercomputers
By Eric J. Lerner
| |
Today's computer processors are ten times faster than those of five years ago, so why don’t computers perform at one-tenth the time or their predecessors? A major bottleneck is computer memory. Typically, a PC's random access memory (RAM), contains only one percent of the data stored on the hard disk drive. But a hard drive supplies data nearly a thousand times slower than a processor can use it. So most of the time, a processor is waiting for data to be found and transferred.
This problem is only getting worse as the already massive amount of digital data—everything from customer information to genetic sequences—continues to increase. "What's needed today is not so much the ability to process each piece of data a great deal," explains Mark Dean, IBM vice president of Systems, "it's the ability to swiftly sort through a huge amount of data."
Cellular architecture from the ground up
Early in 2000, a group of IBM researchers and engineers began to solve this problem using a future generation of ultra-high performance computers. The group proposed a new computer design, called cellular architecture. "The key idea is that instead of focusing on processor micro-architecture and structure, as in the past, we optimize the memory system's latency and throughput—how fast we can access, search and move data," explains Dean.
Cellular computers, like existing supercomputers, consist of many identical processors. However, cellular computers differ from supercomputers in a number of important ways, each designed to speed memory access. "A cellular machine has memory units integrated into each of its cells, instead of having a central bank accessed by all the processors," points out George Chiu, one of the scientists working on IBM's effort. This means that processors don't have to take turns waiting for access to central memory. In fact, in one of IBM's experimental cellular machines, Blue Gene, each processor will be fed data by eight different memory sub-sectors, each connected with a different subtask or thread. "It's sort of like having a chicken run from one feed spigot to another so the chicken is always feeding," Chiu says.
Second, when data is needed, all cells can be used to look for it in their own memories, instead of having a central controller deciding which processors should do the job. "This means that some processors will waste time looking for data that is not in their cells' memory, but the data will be found more quickly this way," says Dean.
For this architecture to work on very large databases there must be a high number of cells, with tens of thousands to a million processors, rather than the hundreds of processors found in existing supercomputers. Also, unlike existing supercomputers, each cell is small enough—a single chip—to enable this extremely large-scale parallel operation (the ability to divvy up instructions and perform them simultaneously). To shorten communication time between cells, each is connected only to its nearest neighbors.
Finally, with so many processors, some failures are inevitable, so cellular architecture automatically reroutes instructions and data around malfunctioning cells.
"Cellular architecture uses ideas that have already been developed, such as putting a whole system—memory and processors—on a chip, and using many processors in parallel to carry out instructions, but pushes them to extremes to get the fastest data acquisition and search possible," says Dean.
Folding protein with Blue Gene
To explore the technology needed for cellular architectures, IBM is building three large- scale machines: Super Dense Server, a sister project to Blue Gene and Blue Gene. The Super Dense Server, which will be built with off-the-shelf components, represents an evolutionary step beyond existing high-end servers. By using smaller processing elements, IBM researchers expect to be able to squeeze 500 to 600 nodes in a single 19-inch rack, which is about five times more than current commercial Web servers can accommodate. "While the Super Dense Server is not as dense as our other cellular machines, it will help us to develop the software required for using very large numbers of nodes," Dean explains.
A sister project to Blue Gene is a more general-purpose supercomputer with a large memory capacity, which IBM is developing to handle very large databases and to demonstrate basic cellular design. With 64,000 to 80,000 processors, this machine will be 15 to 20 times faster than any existing computer, running about 180 to 256 teraflops (trillion calculations per second). But its key feature will be its considerable fast memory. With each cell boasting 128 to 512 megabytes of RAM, this machine will have up to 40 terabytes of memory. This is the data contained in 5 million full-color photos, or in 6 billion pages of text — one for each human being in the world. Built with modified versions of Power PC processors, the supercomputer is expected to debut in 2004.
Blue Gene, also scheduled for completion in 2004, will be even faster. Unlike the more general-purpose machine, Blue Gene itself will not handle large databases. However, it will give IBM researchers invaluable experience in working with extremely large numbers of cells and in developing special-purpose chips that combine memory and processors in the most efficient manner. With one million processors, each performing a billion operations per second, Blue Gene will be capable of one petaflop—a thousand trillion calculations per second. Such speed will be devoted to a single set of problems—simulating how proteins fold after they are generated in cells.
"If biochemists can see how a protein folds, what the process is, then they can see how to change individual amino acids in the protein. They can then understand how to change the way the protein folds and consequently, its shape. This will change how it acts," Chiu points out. Such changes in function can be vital in developing new drugs. Blue Gene will be able to model medium-sized proteins, but each simulation will take a full year even with the computer's immense speed.
All of these machines will be one-of-a-kind or few-of-a-kind models, not prototypes for any specific products. In the past, such one-of-a-kind projects have paved the way forreal products. "We developed GF-11 to do one problem in particle physics," says Bill Pulleyblank, the new director of mathematical sciences and of the IBM Deep Computing Institute, "but with what we learned we were able to design the SP-2 supercomputer series. So I think we'll get equally good results from the cellular architecture projects." IBM is betting about $100 million on the construction of Blue Gene alone, an indication of researchers’ confidence that commercial applications will be available in the near future.
