by Tim Staub, associate editor LIVEwire
Dallas, Texas — The following is an interview with Jan Silverman, vice president, SGI High-End Servers & Graphics:
HPCwire: SGI’s main message at SC2000 seems to be “modular computing.” What does that actually mean?
SILVERMAN: One thing SGI realized years ago is that supercomputing means different things to different people. Some people need large processor counts. Some need tons of data or bandwidth. Some need large, addressable shared-memory spaces. Some have complex visualization problems. Our goal has been to build a computer that can meet all of those needs in a scalable fashion, so users can tailor systems to the type of supercomputing capabilities that their specific applications require. Our SGI Origin 3000 and SGI Onyx 3000 series machines are built from component modules-we call them bricks-and the customer can decide how many bricks of each type to put in a system. We call this design approach NUMAflex.
HPCwire: Why did you choose the term “NUMAflex”?
SILVERMAN: We call it NUMAflex because the modular approach is enabled by our third-generation SGI NUMA architecture. Our main focus for differentiation in the supercomputing space has been on architecture. As different technologies emerge, we find customers who want to stay up to date with the latest processors, the latest I/O standards, the highest bandwidths, and the best in visualization. With a NUMAflex modular approach, you can stay up to date with all the latest technologies. Changing one piece of technology, such as an I/O standard or a faster system interconnect, doesn’t mean you have to upgrade the rest of the system. This is a level of investment protection that the industry hasn’t seen before and a big reason why modular computing has emerged as a superior design to backplane SMP systems.
HPCwire: How does the performance compare between SGI NUMA and traditional SMP?
SILVERMAN: What really got us started on NUMA in the first place is scalability. Our customers are solving big problems, and they need big computing. But in a traditional SMP system, the more CPUs or I/O cards you add to the system, the more traffic you create on the midplane or backplane. It’s like adding more cars to the freeway. Everything slows down. SGI NUMA gives each processor multiple paths for accessing remote system resources at extremely low latencies. That’s what broke the hardware barrier to scalability and let us build much larger shared-memory systems-up to 512 processors-while the rest of the industry was stuck at 64 CPUs or less. More importantly, it gave us real application scalability. The more processing power, memory, or bandwidth you added to the machine, the faster your application would go. It’s real application performance, not theoretical peak numbers that can’t be achieved.
HPCwire: How do you manage these very large systems?
SILVERMAN: Again, the modularity really helps us here. The key is that the biggest systems are built out of exactly the same bricks, or modules, as the smaller ones. You just add more bricks to make the system bigger, and you use the same operating system and application binaries across the whole space. So the big systems get administered exactly the same as the smaller ones. The machine is also a lot more resilient as a result. Think about your modular home entertainment system, which is a great analogy. Let’s say your VCR breaks. That doesn’t mean you don’t want to watch TV or listen to CDs. Because it’s modular, you might lose a component or part of the capability until it gets fixed, but you don’t lose the whole system. The modularity makes the whole machine easier to service and gives you better total system uptime. And if you wanted to replace your VCR with a DVD player, you’d still want to keep your television, receiver, and speakers, right? You upgrade each of the modular components separately to stay on top of the latest technology, but you protect your investment on the rest of the system.
HPCwire: So what’s the next big thing from SGI going to be?
SILVERMAN: SGI is totally committed to supercomputing, and our technical and creative customers are always pushing the envelope of technology to solve problems in a way that has a big impact on the world. In a way, their innovations drive our innovations. We want to continue to deliver the best solutions in big compute, big data, big bandwidth, and big visualization, which keeps our customers ahead of their competition. But the key difference with our NUMAflex modular design is that now we can keep delivering the latest technologies in a way that protects your original investments. Take any of the big new technologies we expect to see in the next year, such as Intel IA-64 Itanium processor, higher-bandwidth PCI-X peripherals, or faster system interconnects or bigger, faster disks. At any point, you can incorporate any of these technologies into your architecture at a modular level, protect your investment on the rest of the system, and achieve true application scalability. That’s a critical, revolutionary innovation that only SGI can deliver.