July 9, 2014

TSUBAME Prototype System Balances Benchmark Leadership

Nicole Hemsoth

When it comes to large-scale supercomputing installations, Asia is the continent to watch carefully over the next few years. Already host to the top system in the world, China’s Tianhe-2, Japan and others have ambitions to take over the top ten list of systems.

In the most recent rankings, Japan is home to 30 systems out of the Top 500 worldwide supercomputer share, up from just 18 systems in 2010. In addition to the #4 K Computer system at RIKEN, another noteworthy machine, the #14-ranked TSUBAME 2.5 (or TSUBAME KFC, named for the Kepler Fluid Cooling component, which ironically means it’s submerged in oil) is garnering attention. Part of why this 76,032-core system is one to watch is because it’s setting the stage for a new machine with far higher performance—and some novel integrations of unique storage, network, middleware and other technology.

TSUBAMEKFCsmallThis month at ISC, Satoshi Matsuoka from the Tokyo Institute of Technology presented an overview of progress with the TSUBAME KFC machine, which is the precursor and prototype for the next-generation system of the same name that the team will roll out sometime in 2016. When it emerges in 2016, TSUBAME 3 is expected to hit the 25-30 petaflop performance mark while balancing some new technologies cooked into the middleware, storage, and network. In addition to the “Kepler Fluid Cooling” which was at the heart of its top efficiency rankings on the Green 500 this summer, where it was the top system.

“Assuming that TSUBAME-KFC’s energy efficiency could be scaled linearly to an exaflop supercomputing system, one that can perform one trillion floating-point operations per second, such a system would consume on the order of 225 megawatts (MW),“ said Wu Feng of the Green 500. “Although this 225-megawatt power envelope is still quite far from DARPA’s optimistic target of a 67-megawatt power envelope, it is an order of magnitude better than the initial projection of a nearly 3000-megawatt power envelope from 2007 when the first official Green500 list was launched.”

But the Green 500 and prototype system’s placement in the Top 500 are just part of a larger story–one that Masuoka doesn’t want the community to overlook. It’s about handling the next generation of data-intensive applications, which is an area full of lessons from outside of supercomputing.

The focus of TSUBAME-3 (and leading into 4 in the 2020-20222 timeframe) will be on balancing efficiency, data-readiness, and of course performance or, as Matsuoka described in his talk, a convergence of supercomputing with extreme big data. We are all aware of the bubble big data has presented in HPC, but Mastuoka says it’s critical to design systems that integrate lessons learned from hyperscale cloud datacenters as well as what appears ahead for eventual exascale-class systems.


The current KFC machine ranked #12 on the Graph 500, and #6 on the Green Graph 500, which looks at the energy efficiency of solving “big data” graph problems. This is where the real future focus of the system in its 2016 incarnation will be, says Matsuoka. As he explained, at the beginning of the Graph 500 list, the expectation among some was that the list would look far different than the Top 500 with a number of cloud vendors submitting their distributed machines for the rankings. However, the list looks quite similar to the Top 500, with the same machines at the top of the list that are in the Top 500 and to a lesser extent, the Green 500. The hope is to balance top results across these categories with an eye on real-world applications, not just benchmark toppling.

These early predictions stood to reason since ostensibly, the big clouds were tackling “big data” jobs The common estimate is that a giant web services company like Amazon has around 500,000 nodes with around 6 million cores spread throughout its network. That makes for a massive distributed machine, but the core counts of these cheaper servers are often far lower than ultra-dense supercomputers. For instance, Tianhe-2 has 3 million cores spread across 18,000 nodes. Matsuooka says this point isn’t a surprising one—large datacenters are common, but they tend to be very sparse; they don’t require the networking and density of supercomputers—and therefore don’t have the same capability.

The goal of the next incarnation of TSUBAME in 2016 will be reducing the size of the system while supplying the needed bandwidth and compute horsepower in a much smaller amount of space. Cheap SSDs, ultra-dense system design, and leveraging new uses of burst buffer technology to offload critical processing tasks are key to the approach with both the coing TSUBAME 3 and the future 4 machine.

More text here

TSUBAMEKFCMatsuoka says that TSUBAME 3 will feature larger capacity SSDs that will give the Tokyo Tech team local bandwidth of about 50 TB of capacity, or 50 GB/s bandwidth in a single rack, suppose 40 racks, several terabytes per second of aggregate bandwidth. They’re working with DDN now to further this future.

The Top 500 list in 2016 is set to be an interesting one, particularly in November, with the addition of this next-generation machine and several others we’ve heard word of. While not all the major machines set to come online by Linpack time will be running the famous benchmark since it doesn’t adequately reflect their goals, Japan is expected to take advantage of all three major benchmarking opportunities–Top 500, Green 500, and Graph 500–to show the balanced system they’re seeking…one that’s ultra-efficient, big data capable, and of course, high performing in a top 10-class way.

Tags: , ,