In May, Sun Microsystems, Inc. announced that the Tokyo Institute of Technology (Tokyo Tech), one of the world's premier science and technology universities, installed a 38-teraflops supercomputer, based on Sun Fire Server technology. Tokyo Tech's supercomputer, called TSUBAME, represents Sun's largest high performance computing win to date. The system also represents the largest supercomputer outside the United States. It includes 10,480 AMD Opteron processor cores, more than 21 terabytes of memory and 1.1 petabytes of hard disk storage.
IDC's Earl Joseph recently completed an exclusive interview for HPCwire with Professor Satoshi Matsuoka, of the Tokyo Tech's Global Scientific Information and Computing Center. Professor Matsuoka talks about the significance of the new TSUBAME system and describes some of the cutting-edge applications that will be running on it.
HPCwire: What's the status of the Sun system?
Matsuoka: The Sun system was installed in March 2006. When the Sun boxes arrived in mass, the infrastructure was already in place, so the installation went quickly, one of the fastest installations ever of such a large system. On April 3, we were able to start the service on a limited number of nodes. We did preliminary Linpack runs on the whole system in early May. In early June, about 5,000 processors, or half of the system, was opened up for public use. We were reserving the other 5,000 processors for benchmarking and special allocations.
The user workload has been growing steadily and is already at 70 percent. We're now running hundreds of jobs on the system. We're receiving help desk mail from all of the university departments we serve. Many users are telling us they love the new system. Some users are starting to do massive runs. One of these, for example, is looking at scenarios with avian flu and the impact of potential social policy decisions. This is a parameter study.
HPCwire: How difficult was the installation?
Matsuoka: Based on our prior experiences bringing up large computers, the glitches we've faced with this Sun system have been quite minor. We're currently stress-testing the machine around the clock. There have been no major complaints. It's just been an amazingly quick ramp-up to stable operation. After just a few weeks, failures were minor and very sporadic.
HPCwire: Were there any surprises?
Matsuoka: There are always a couple. Staying with fat nodes turned out to be a good strategic move. With a small node system of this size, we would have had to manage 10,000 thin nodes. With fat nodes, the number is only 655, which is not all that different from the scale we have been used to. The other surprise is how much attention we are getting because of this system.
HPCwire: I've been hearing a lot lately about the issue of power consumption on large-scale HPC systems. How has your new system affected power consumption at your center?
Matsuoka: With our older systems, which included a vector supercomputer and others, we were operating at 600-700 kilowatts for a total of two peak teraflops. Now we've raised the peak 40 times to 80 teraflops, and power consumption has increased by a factor of less than two. It's at 1.2 megawatts today, but we're implementing power-saving measures that we believe will lower this 10-15 percent to about 800 kilowatts for the whole machine. We're getting a lot more accomplished for this modest increase in power consumption, and this is more than offset by the benefits the new system brings to the university, including the ability to attract more research funding.
HPCwire: Did you have to build or find more floor space?
Matsuoka: No. The Sun design is amazingly compact. All 76 racks and 20 cooling units fit into our existing building.
HPCwire: Are any end-users doing work on the ClearSpeed processors yet?
Matsuoka: To some extent, but it's still very early. We envision people running smaller applications and using their libraries to exploit the ClearSpeed accelerators. For the right applications, there will be big performance gains.
HPCwire: Where do the codes come from that run on the new system?
Matsuoka: We help people from the university port and scale their codes. We have some user accounts from outside Japan, and we get outside codes to run, such as the Dynamo and ocean codes from the Earth Simulator. It's a big variety. We're running a combustion code on the system, for example, and we expect to run a carbon nanotubes code because the university has a well-known expert in this field. We also have civil engineering codes. We're getting codes from every university department that uses HPC.
HPCwire: Why would some Earth Simulator codes be run on your system instead of on the Earth Simulator itself?
Matsuoka: The Earth Simulator is already being used to capacity. We are providing room for growth.
HPCwire: How hard is it for Japanese users who aren't part of the university to access your system?
Matsuoka: Today, they need to have a collaboration agreement with the university, and agree to make the results available to the public. In the future, we will have more allocations for the other national centers. We still need to develop chargeback methods for this. Also, it would help if the Japanese government could establish funding methods to more easily allow the national centers to share resources over the Grid.
HPCwire: How will you provide access over the Grid?
Matsuoka: Users will have national Grid accounts with allocations. We are in the process of establishing such a national CyberScience Infrastructure Grid starting this year.
HPCwire: What about job sizes? What do you expect to happen there?
Matsuoka: On our older machines, we saw MPI job sizes ranging from one to several hundred processors. Once you have a larger machine, people want to take advantage of it. We expect to run some codes that use a large fraction of the whole system. Then there will be more codes that run on 64 or so processors and are run many times. These are parametric studies for ensembles, like the bird flu example I mentioned earlier. We've already run benchmarks using half, and in some cases all, of the machine.
HPCwire: Will you run international benchmarks?
Matsuoka: Yes. The process is that we would first need to sign a contract for collaboration with them that includes making the results public and posting them on our website.
HPCwire: What programming languages are being used on the Sun system?
Matsuoka: People use MPI, OpenMPI, OpenMP, Java, MATLAB, scripting language, automatic parallelizing compilers, and so on. We may look into UPC once the system is even more stable.
HPCwire: How would you handle 1,000 small accounts accessing the system at once?
Matsuoka: That would be evidence of major success. Demand like that would demonstrate that we needed more resources and capacity in future upgrades and systems.
HPCwire: How hard would it be to upgrade the system?
Matsuoka: The Sun design makes it easy to upgrade to quad-core AMD processors. You can replace parts of the system or the whole system. We couldn't do this with our previous systems.
HPCwire: How would you sum up the experience so far?
Matsuoka: It's an absolutely great fit with our mission, which is to be an incubator for research within our educational institution; to provide an HPC resource that is large, easily accessible and easy to use. These are the biggest wins for us. The Sun system also supports our objective to develop HPC users who can run moderate size problems today, and then later can move easily to even larger HPC systems, including Japan's planned 10-petaflops system.