by Steven Witucki, assistant editor LIVEwire
Dallas, Texas — On August 22, 2000 the U.S. Department of Energy’s (DOE) National Nuclear Security Administration (NNSA) selected Compaq to build the world’s fastest and most powerful supercomputer, a 30+ TeraOPS system code-named ‘Q’-the latest advancement in the Accelerated Strategic Computing Initiative (ASCI).
Richard Kaufmann, Technical Director for Compaq Computer Corporation’s High Performance Technical Computing Group, recently took the time to speak with HPCwire about Q and the state of high performance computing at Compaq.
HPCwire: Tell us about some of the challenges of developing a system to meet the ASCI requirements for the 30+ TeraOPS Q supercomputer.
KAUFMANN: The Q system will be the largest computer in the world when installed. It will push on all system parameters: scheduling, I/O, system management, MTBF, etc. It will be based on our AlphaServer SC technology. SC systems are installed at CEA (France), LANL, LLNL, ORNL, Pittsburgh Supercomputer Center, and others.
Compaq’s strategy (with guidance from Los Alamos) is to build Q out of approximately 375 of Compaq’s largest servers, the 32-CPU AlphaServer GS320. The GS320 is shipping to customers now, and has proven to be a very stable platform. The Q will use some planned upgrades for the GS320, available late 2001. The upgrades include faster CPUs and an updated I/O subsystem. Hardware stability is key to a successful deployment of Q, so using a solid server is very important.
Q will use the next version of the interconnect fabric from our AlphaServer SC series. The main changes are a different, faster host interface and increased network link bandwidth. Q will have eight rails (each rail is a separate network plane), and each GS320 will have at least 4 GB/s of message passing bandwidth.
375 nodes is a lot of nodes! Unless, of course, you compare Q with the NSF TCS-1 system to be installed at the Pittsburgh Supercomputing Center in 2001. Their system will consist of more than 600 quad-processor servers, and will be built with two rails of the AlphaServer SC fabric.
Q has a strong I/O requirement. An application must be able to dump all 12TB of memory to a global file store in ten minutes. This translates to more than 20GB/s of sustained parallel file system write bandwidth! We’re using a request forwarding mechanism to move I/O from the compute nodes to the file servers. This forwarding technology will first appear in the next few months in AlphaServer SC version 2.0, and we expect to spend significant time on this mechanism to ensure proper scaling.
We’re fortunate to have an incredibly solid compiler. It is easy to forget that modern processors are totally dependent on very smart compilers (not to mention compiler writers!), and Compaq’s Gem compiler technology has been one of our secret weapons. Some new technology will be added in Q’s timeframe to help with “NUMA” memory topologies. There will be a sneak peak of this technology during Jonathan Harris’ talk at Supercomputing.
There are nearly 12,000 Alpha processors in Q, and all of them will want to send MPI messages to each other. The virtual DMA technology of the AlphaServer SC fabric makes this practical, but we expect to spend a significant amount of time tackling application scaling. This is one of the key areas where the researchers at LANL will be working with us.
There are many other challenges in building Q! Just imagine this huge number of servers, each of which is six feet wide and five feet tall. There are more than 3,000 parallel fiber links from the servers to the switches, and an equal number of copper links tying together the cabinets for each switch. We’re building a special Q outpost in our manufacturing plant just to handle the first stage of system integration. The logistics required for this effort are pretty impressive.
HPCwire: Tell us about Compaq’s involvement with the Accelerated Strategic Computing Initiative (ASCI).
KAUFMANN: Compaq (then DEC) has worked with ASCI since the ASCI Blue procurement in 1995. Our first successful procurement was ASCI PathForward. Under the PathForward program, Jim Tomkins (Sandia), Karl-Heinz Winkler (LANL) and Mark Seager (LLNL) worked with us to accelerate our interconnect program. PathForward is directly responsible for our relationship with Quadrics Supercomputers World, the supplier of the interconnect fabric in AlphaServer SC, and we wouldn’t have been able to bid on the 30T system without it.
Our discussions with the DoE labs over the past five years have profoundly influenced every aspect of our system design. Personnel from the DoE labs (as well as some other advanced customers, such as the French CEA) are involved in our design decisions right at the point of conception; quite often these researchers know about our future designs well before many of our own engineers!
HPCwire: What can you tell us about how the National Nuclear Security Administration (NNSA) will use Q?
KAUFMANN: Q will be used to push the state of the art of scientific modeling and simulation. The DoE will use the system to provide responsible stockpile stewardship, in an environment where nuclear testing is no longer acceptable. This need for reliable simulation is pushing not just computer companies, but also legions of algorithm and code designers – at the DoE, universities, and ISVs.
HPCwire: Does Compaq plan to continue to support governmental agencies like NNSA? If so, which agencies?
KAUFMANN: Compaq has long-standing relationships with the NSF, the intelligence community, the DoD, NASA, and others. Our HPTC organization works closely with these U.S. agencies, as well as other organizations outside the U.S.
If anything, we’re looking for ways to strengthen our relationships with these agencies. We call these accounts “lighthouse accounts” because we expect their needs to be a few years ahead of the wider market. It is working to satisfy the needs of these accounts that produces our technology for future generations of AlphaServer SC.
HPCwire: Will Compaq be building more 30+ TeraOPS systems, or will they be focussing on creating yet more powerful systems?
KAUFMANN: We’re deeply involved in design discussions for 100TF and beyond. However, we do expect to sell lots systems in the 0.002TF (aka a single CPU!) to the 30 TF range over the next few years. Just go to http://www.compaq.com/hpc , and have your credit card ready!
HPCwire: Compaq recently reported record revenue for its third quarter this year. Do you think that the selection of Q by the NNSA contributed to this?
KAUFMANN: It’s our belief that the Q announcement (as well as some other major successes, described later) has sent a clear message that Compaq is in an incredibly strong position in HPTC. This undoubtedly has helped us in recent competitive situations, but significant revenue for the Q system itself won’t hit Compaq until 2001 and 2002. It’s great to work for a company that’s firing on all cylinders. Everything from the handheld iPaq to the 32-CPU AlphaServer GS320 are doing quite well, and the HPTC group is proud to do its part to help Compaq’s financial success.
HPCwire: Is there any other news regarding High performance computing at Compaq that you would like to talk about?
KAUFMANN: This has been a great year for us! In addition to Q, we’ve won the largest supercomputer program in Europe, CEA, the largest civilian supercomputer, the NSF Terascale system with Pittsburgh Supercomputing Center, and very recently, the largest supercomputer in Japan, with the Japanese Atomic Research Institute. Our AlphaServers played a pivotal role in mapping the human genome earlier this year and at this moment are allowing researchers at Celera Genomics and many publicly funded institutions to complete the annotation of the genome and publish their findings. Blue Sky Studios, part of Fox, selected us to deliver the most powerful computing facility in the entertainment industry last spring. And in May, we began shipping our new GS series servers; high performance computing customers have been snapping them up at a record pace. So we’re on a very strong roll.
Looking forward, Compaq’s roadmap would be quite daunting to deliver, except for the excellent partners we’ve picked: Quadrics’ Elan network is the fabric of the AlphaServer SC series. This fabric has given us a very strong performance and capacity boost. We’re also lucky to be working with Etnus (TotalView), Pallas (Vampir SC), Platform Software (LSF), and Raytheon (integration help on Q).
By the way, if anyone out there would like to help, Compaq is busy hiring talented folks for its High Performance Technical Computing organization. Feel free to come and talk to us at Supercomputing!