Home > Inperso > Full Story

Powering processors for Unix

Five years ago, IBM set out to build a system that would change the way one looked at conventional Unix servers. The result is the pSeries 690 server (code named Regatta) which was launched in October. With a new processor called POWER4 and features borrowed from its mainframe systems, Regatta might just be a threat to Sun Microsystems, which has for long been dominant in the server market. Dr. Joel Tendler, Program Director, Technology Assessment, Enterprise Systems Group, IBM has been one of the architects of this system. In an exclusive interview with Network Magazine he speaks about the innovations that go into Regatta and explains how it stands up against the competition. by Brian Pereira

You have been instrumental in designing the POWER4 processor. What are some of the bottlenecks and technical limitations with processor architecture? How has IBM worked round these bottlenecks during development of the POWER4 processor?
When we began working on POWER4 we asked ourselves, how do you get the most out of the transistors you have? How do you build those transistors? If you look at our development over the last few years, we have been announcing breakthroughs almost on an annual basis. In 1997 we announced copper (as a conducting material on the chip). Up to that time chips were using aluminum which generates too much heat. Copper is a better conductor and hence the chip runs cooler and faster. The next breakthrough was Silicon on Insulator (SOI) where we put a protective layer inside the silicon, effectively giving the electrons a shorter path between gates. Both of these innovations are inside our POWER4 chip. We have also announced other innovations like 'low-K' dielectric which reduces crosstalk enabling us to move the copper wires closer. But it's not just having good chips that matters—it's also about using that technology. Technology goes beyond silicon—what matters is how you build systems and put them together.

As you increase the number of processors within a system the number of interactions (between components) goes up. That's where the Switch comes in. Until now, servers used Crossbar Switches. But the circuitry necessary to make this work increases (as you add more processors). With POWER4 we use a Distributed Switch which does not have a single point of control—the control is spread out among the chips. With this switch I can pass information between my processors with higher bandwidth and in shorter time.

The bus is something we call Synchronous Way Pipeline Interface. It allows me to run the buses at speeds never achieved before. With the POWER4 system the bus speed between chips is 650 MHz. Sun's Ultrasparc III processor communicates at 150 MHz, so our bus speed is four times as fast. These are dual buses that run at half the processor speed, and that will continue to be the case as we increase processor speeds in future.

IBM has been building mainframes for 30 years and has a good knowledge of that technology. It is now putting that technology in Unix/midrange systems. What kind of mainframe technology goes into the new eServer pSeries 690?
We have put in things like Chip Kill technology, partitioning and self-healing. Chip Kill technology says that even if a memory chip dies/fails I can reconstruct all the information inside that failed chip. This increases the reliability of the system's memory.

Partitioning technology allows you to virtualize servers. Our p690 system allows you to build up to 16 logical partitions, but it's not just the number—it's how you build those partitions. The resources in these partitions are uncoupled and each partition can be configured to have a specific number of resources. The granularity of these partitions is one processor, one PCI adapter and a minimum of 1 GB memory, in increments of 256 MB.

Our competition (Sun) offers 18 domains with 72 processors, but those are physical partitions. Partitioning was first introduced on the IBM mainframes in the late 1980s.

What are the migration issues involved when companies move from a room full of servers to a single server with multiple processors?

The application has to be ported, but that's not an issue. There are many customers who have made that migration via AIX. It's more an organizational issue (a decision to move to a single box). We have tools that offer migration from Solaris to AIX. It becomes a lot easier because now I am able to take the multiple boxes and manage them in one place.

That adds value to the customer not just in terms of flexibility but also in terms of TCO. It's not just the box; it's the management of the box, the reliability of the box. It's things like eLiza (self-healing technology)—the vendor can now stop worrying about the box and invest his resource in building applications so that he can satisfy his customers. eLiza technology automatically takes corrective action and thus reduces TCO.

What are the potential applications for the p690 server?
The p690 is designed for both technical and commercial applications. On the commercial side, it can be used by the Banking or Finance, and Manufacturing sectors. These industries demand high reliability, the ability to maintain applications, high bandwidth connectivity, and predictable performance.

The other areas are high-performance computing where it can be used in universities, for weather forecasting, and in the defense industry. I can build a supercomputer by connecting a series of such machines. Each processor is capable of four floating-point operations in a cycle. That's over 5 gigaflops. With nine floating-point operations per processor, that's over a 166 gigaflops for a 32-way system. Take six of those systems and I have a teraflop.

What is the basic configuration of the p690? What is its scalability like?
The p690 comes in modules. I can have a 1-, 2-, 3-, or 4-module system. Each module has 8 processors. So it can be as small as an 8-way symmetric multiprocessor system to as large as a 32-way symmetric multiprocessor. So we offer an 8-, 16-, 24-, or 32-way systems. The processors are either 1.1 GHz (1100MHz) or 1.3 MHz. In future we will move to 1.5 GHz and 2 GHz, and the buses will scale with it.

We believe that you must design today not just with the technology that is available, but with the technology that will come in the next few years.

How does the price of this machine compare with that of the Sun Fire 15K? Assuming both have similar configurations.

Regatta is half the price of a Sun Fire. On a processor-to-processor base, we offer twice the computing power at half the cost. Our 32-way system outperforms Sun's 72-way server.

So you are looking at putting fewer processors in the box?
We are not looking at putting fewer processors in the box, but putting in enough processors to do the job.

Fewer processors add value to the customer too. Because many software vendors charge by number of processors. If we offer 32 processors at 1.3 GHz it is a much smaller number than 72 processors on 900 MHz (as in the Sun Fire). So I have software savings in addition to higher performance.

On the technology front how does Regatta compare with HP Superdome and Sun Fire?
Both of those are using technology that is a generation or two old. We started this project five years ago (end of 1996), and when we began we tried to forecast where our competitors would be now and set our targets for the Regatta system. Both HP and Sun have slipped years from where they said they would be. We delivered Regatta on the date we said we would five years ago.

Their switches are a generation behind us. In the area of the chip itself, at 1100 MHz the POWER4 chip consumes 115 watts of power. The Sun at 900 MHz consumes 75 watts. As you increase the frequency, the power increases proportionally, so at 1100 MHz the Sun chip would be 90 watts—that's for one processor. The 115 watts for the POWER4 processor is for both processors and the second level cache.

So it's not just important to have good technology, but also to know how to use it. It's how you build systems using technology. Our chips run cooler. The hotter the chip the lower the reliability. It reduces your TCO.

Comparing our system with HP Superdome, I'd say Regatta outperforms the HP machine and takes up less space. If you look at the HP machine in terms of weight and packaging, there is a significant differences.

Until now developers would first make applications for the Solaris platform. Having a Unix server with new technology does not give you an advantage. What about the applications?

We have set up porting centers where we work with the developers to help them port their applications to our platform. So eventually all the important applications will be there on our system.

Brian Pereira can be reached at brianp@networkmagazineindia.com

<<