HPCwire: Number Crunching, Data Crunching and Energy Efficiency: the HPC Hat Trick

Since 1986 - Covering the Fastest Computers
in the World and the People Who Run Them

Translation Disclaimer

Subscribe | Sign In

Visit additional Tabor Communication Publications

Leading HPC
Solution Providers

February 02, 2012

Number Crunching, Data Crunching and Energy Efficiency: the HPC Hat Trick

Gary Johnson, Founder, Computational Science Solutions

In the world of high performance computing, there are three distinct metrics in play: number crunching speed; data crunching speed; and energy efficiency. Can a computer excel at all three, or is our best recourse to try for something less than a hat trick?

An abundance of metrics

In the past, number crunching ruled HPC. The measure was LINPACK, the metric was FLOPS (floating point operations per second) and the list was the TOP500. Currently, data crunching has soared in importance and visibility and is arguably on par with number crunching. The measure here is an evolving set of kernels from graph algorithms, the metric is TEPS (traversed edges per second) and the list is the Graph 500.

Simultaneously, the previously unconstrained race to the top is being supplanted by a new form of competition – one constrained by electrical power. Hence the Green500 list. Here, the measure is energy efficiency and the metric is MFLOPS/watt.

The explicit constraint, introduced as a design goal for an exaflop computer, is one exaflop/20 MW or 50 gigaflops per watt. The computer currently at the top of the Green500 list operates at slightly over two gigaflops per watt. So, the 20 MW design goal is quite ambitious. However, it has now brought energy efficiency to the forefront of HPC. It’s not just for tree huggers anymore.

Loneliness at the top

The top of each of these lists is a lonely place and an expensive one to inhabit. Most machines will never attain it, but this is the realm where a lot of innovation happens. If only a few machines can reach the summit, then one probably can’t afford to have them be highly effective on only a small set of applications. The cost is simply too great. Other opportunities for research investments would probably outcompete a narrowly targeted machine.

So it seems that at the top end of HPC, we are now seeking energy efficient computers that perform well at both number and data crunching. Can we get all three? Let’s take a look at the current state of the art.

Graph-Top-Green 500 list comparisons

The current version of the TOP500 list cites energy efficiencies for a large number of its entries, while the Green500 list provides energy efficiencies for all of its entries. Meanwhile, the Graph 500 list is still very much a work in progress. The current version contains only 49 distinct computers and does not provide any mapping of these to either the TOP500 or the Green500 list. Nonetheless, it is possible to locate at least 19 Graph 500 computers on the other two lists. So we can make at least a partial comparison of these lists, the results of which are shown below.

The machine nicknames used are self-explanatory, except possibly for “NSQP2” which refers to the NNSA-Office of Science BlueGene/Q Prototype II. The machines are listed in the order of their average ranking across all three lists.

Some obvious conclusions may be drawn:

Tsubame is the clear overall winner.
A couple of machines show a reasonable balance over all three metrics – Tsubame and Gordon
If one picks Top & Green, Tsubame and Gordon are still the best.
For Graph & Green, NSQP2 wins, Tsubame is a close second, and Endeavor-W, Endeavor-S and Gordon look pretty good.
For the choice Graph & Top, there are surprisingly many good choices, including: Tsubame, Hopper, Intrepid, Jaguar, Jaguar PF, Kraken, Kraken-F, Lomonosov, Franklin, Lonestar and Red Sky.

The top five of the TOP500

Note that three of the top five TOP500 computers are missing from the comparison: K Computer (#1); Tianhe-1A (#2); and Nebulae (#4). This is because they’re not currently included in the Graph 500 list. Jaguar (#3) and Tsubame (#5) are present. If the missing machines are added, the comparison looks like this:

The only comparison now possible is Top & Green. As is shown above, Tsubame still wins. However the three new entries all score higher than Gordon.

Can one machine have it all?

Based on the limited sampling used here, it appears that the answer is yes. Tsubame is leading the way. Kudos to Prof. Satoshi Matsuoka and his team at the Tokyo Institute of Technology’s Global Scientific Information and Computing Center (GSIC). It will be interesting to see if this conclusion remains true as we move along the path to exascale.

Secret sauce

Those curious about Tsubame’s secret sauce may consult the Tsubame2 System Architecture information on the GSIC website. Here’s a brief extract:

TSUBAME2 is a production supercomputer operated by Global Scientific Information and Computing Center (GSIC), Tokyo Institute of Technology in cooperation with our industrial partners, including NEC, HP, NVIDIA, Microsoft, Voltaire among others. Since Fall 2010, it has been one of the fastest and greenest supercomputers in the world, boasting 2.4 PFlops peak performance by aggressive GPU acceleration, which allows scientists to enjoy significantly faster, larger computing than ever. This is the second instantiation of our TSUBAME-series supercomputers with the first being, as you might guess, TSUBAME1. It also employed various cutting-edge HPC acceleration technologies, such as ClearSpeed and NVIDIA GPUs, where we had learned many important technical lessons that eventually played a crucial role in designing and constructing our latest supercomputer. Compared to its predecessor, TSUBAME2, while keeping its power consumption nearly the same as before, achieves 30x performance boost by inheriting and further enhancing the successful architectural designs.

Key architectural points cited are:

Extended usage of GPU accelerators
Much improved intra- and inter-node bandwidths
Petascale high-bandwidth shared storage
Ultra-fast local storage (SSD)

What about number Crunching versus data crunching?

Another noteworthy observation from the Graph & Top comparisons is that there are quite a few computers that seem reasonably well balanced for a mix of both number crunching and data crunching tasks. The conventional wisdom is that number crunching and data crunching take advantage of significantly different computer attributes and that a single computer architecture may not work well for both kinds of tasks. The limited sampling used here appears to contradict that view.

Caveats

As previously mentioned, the Graph 500 list is a work in progress. As it matures and as the list expands to encompass more machines, the conclusions presented here could change.

In the comparisons made here, equal weight has been given to the importance of placement on each of the lists. If one assigns different weights, the conclusions may change. However, it appears that any “reasonable” set of weightings would yield substantially the same conclusions.

-----

About the author

Gary M. Johnson is the founder of Computational Science Solutions, LLC, whose mission is to develop, advocate, and implement solutions for the global computational science and engineering community.

Dr. Johnson specializes in management of high performance computing, applied mathematics, and computational science research activities; advocacy, development, and management of high performance computing centers; development of national science and technology policy; and creation of education and research programs in computational engineering and science.

He has worked in Academia, Industry and Government. He has held full professorships at Colorado State University and George Mason University, been a researcher at United Technologies Research Center, and worked for the Department of Defense, NASA, and the Department of Energy.

He is a graduate of the U.S. Air Force Academy; holds advanced degrees from Caltech and the von Karman Institute; and has a Ph.D. in applied sciences from the University of Brussels.

Share Options

Subscribe

Subscribe to HPCwire

Discussion

There are 3 discussion items posted.

Log plot
Submitted by Michael Perrone on Feb 3, 2012 @ 5:38 AM EST

Thanks for the interesting comparison. Might be nice to see this data on a log plot or truncated after rank ~50 or less. That way we could see the ones that matter more clearly.

Post #1

graphing is not synonymous with data handling
Submitted by Gary on Feb 3, 2012 @ 8:22 AM EST

The Graph benchmark rationale is, in part, "Graphs are a core part of most analytics workloads."

My shop handles lots of data and we use analytics to develop data handling processes that we use in production. But graphing data for analysis is not used in production. We continue to develop analytics and our data handling processes but in development mode. In our HPC production environment we find data handling processing demands are significantly different than our numerical modeling processing demands.

Post #2

a missing consideration
Submitted by Anonymous on Feb 4, 2012 @ 2:43 PM EST

One thing to note is that the Green500 is measuring a ratio (FLOPS/watts) while the Top500 and Graph500 are measuring absolute quantities (FLOPS and TEPS respectively). It is reasonable to assume that FLOPS/watt will scale fairly linearly with machine size and in fact, the Green500 allows you to measure a portion of your system and extrapolate to the full system size. Meanwhile, since the Top500 is measuring peak FLOPS, two machines with similar efficiency (Rmax/Rpeak) may show up in very different places on the list. For example, when BlueGene/P is scaled to its full size, its Rmax will increase, moving it up the Top500 list, while its FLOPS/watt rating should stay relatively constant.

My point here is that while this article compares the characteristics of specific machines like Tsubame, Gordon, etc. it does not provide a a more general comparison of the architectures of these machines, which is typically what someone would consider when making a purchasing decision. Most companies are not going to buy a machine the size of Tsubame or Gordon, though they might want to buy a machine *like* either of these depending on their computational requirements and energy efficiency objectives.

Disclaimer: these comments represent my personal thoughts and opinions, not those of my employer.

Post #3

Join the Discussion