– Pratima H
Sometimes one wonders if it would be easier or tougher for Van Damme if he had to stretch between a bus going forward and the other one going reverse? For quite some time, supercomputers, from the ones carrying the best lineage to those with really-impressive evolutionary DNA, have tried to juggle this balancing act between performance and energy-friendliness.
Why? Does supercomputing muscle really carry some elusive binary equation with power consumption? One can’t help but hope that every next souped-up supercomputer is a splitting image of a greener-than-thou sibling. But more often than not, it’s always a split when it comes down to the brawn on FLOPS (Floating Point Operations Per Second) vs. a greener marrow.
Performance vs. power precision or speed vs. energy exhaust; the tight-rope runs all the way from the world of supercars to supercomputers; and with this year’s latest super-parade we get curious again with these two mutually-orthogonal ladders as we compare, chew and pray.
A Peek Under The Hood
So here it starts. Have a glance at the latest showstoppers on supercomputing runways as per the June 2014 rankings. TSUBAME-KFC – LX 1U-4GPU/104Re-1G Cluster, Intel Xeon E5-2620v2 6C 2.100GHz, Infiniband FDR, NVIDIA K20x made by GSIC Center, Tokyo Institute of Technology managed Mflops/Watt (Mega floating point operations per second, a measure of speed of computers) count of 4,389.82 at a total power of 34.58 KW. This happens to be greenest supercomputer on our planet.
Now swing to the top500 supercomputer pin-up boys on performance and speed and you see that metal-heroes like Milky-way 2 by National Super Computer Center in Guangzhou show Rpeak Tflop/s (Tera or trillions of Floating Point Operations a second) of 54,902.4 at 17,808 KW, and Titan Cray XK7 by Oak Ridge Laboratory hinge at 27,112.5 Tflop/s and 8,209 KW. These are the top two supercomputers in the world right now,
Let’s get the not-so-palatable juxtapositions out of the way first, shall we? There is a name that looks slotted highly enough in the earlier top 500 Green supercomputers list – ROMEO HPC Center – Champagne-Ardenne with 3,130.95 MFLOPS per watt and 81.41 KW clocked on total power. When you try to trace its whereabouts in the general merit order of world’s 500 best supercomputers, it featured after the top 100 supercomputers in that pecking order. The second greenest supercomputer as per that count – Wilkes from Cambridge University was found slotted at 167th in the general top 500 list. It managed 3,631.86 MFLOP/W by the way with a 52.62 Kw mark on total power.
Max-Planck-Gesellschaft MPI/IPP was seen perched at eighth rank in green supercomputing list but it is relegated far south after 200th rank in the general order. TSUBAME-KFC from GSIC Center, Tokyo Institute of Technology with 4,503.17 MFLOPS/W and 27.78 kw in terms of total power is thankfully northbound with a 12th rank in general countdown. And hope is not all lost when we were able to spot the fourth greenest supercomputer Piz Daint from Swiss National Supercomputing Centre (CSCS) yo-yoing between sixth and seventh level (between last year and now) in general list as well
This list saw two petaflop systems (read quadrillions of calculations per second), each capable of computing over one quadrillion operations per second, achieve an efficiency of over 3 gigaflops/watt. One of them was Piz Daint at Swiss National Supercomputing Center and another one was TSUBAME 2.5 at Tokyo Institute of Technology.
It was hard to find parallels or overlaps between the top 10 supercomputers and top 10 green supercomputers, barring the exception of Piz Daint. This gets interesting as the Green500, a commendable effort by the way, which kicked off in April 2005, apparently used the TOP500 list for November to generate the November Green500 list, and the TOP500 June list for the Green500 June list.
From the lens of a science veteran, the distance between the two lists is still substantial since energy efficiency had not been seen as critical an issue as it was going to be. This distance is now expected to reduce and it is now understood that the fastest supercomputers in the coming years are also going to be ranked highly on the green list, argues Prof. Uday Bondhugula, Department of Computer Science and Automation, Indian Institute of Science, Bangalore
Unfortunately or inadvertently, the yardstick used to publicly compare supercomputers to its peers is still the sustained FLOPs on Linpack, although energy efficiency, i.e., sustained FLOPs per watt of power consumed is becoming increasingly significant if one pricks ears to what Prof. Bondhugula adeptly sniffs out.
“The latest X-factor thus is to build supercomputers with accelerators like GPUs and Xeon. This in an attempt to be energy efficient. However, this has its impact on programmability (how easily and productively the supercomputers can be programmed) and their usability for a wide range of applications. Unfortunately, the ranking lists do not capture these criteria, and it is very difficult to devise something which will.”
This is intriguing because every major player who has anything to do in this field- from components, processors, servers, skills, legacy, applications, expertise to clusters; would tell you firsthand how green is as much on the radar as any Linpack metric when one assesses and works towards a better, faster supercomputer.
The workhorses have been munching a lot of green hay, as many will point out.
For Prof Rajat Moona, Director General, C-DAC who was recently basking in a lot of pride having ascended the top 500 list from India and a very good rank in top 10 bracket in Asia-Pacific green supercomputers, decodes that green-efficiency is a formidable upcoming benchmark in supercomputing circles. “How much power is required per compute requirement matters as much as speed of FLOPs. That can be a limiting factor for notching 1000 petaflops. Practical solutions in alignment with power consumed are the buzz of the hour.” He observes how the entire research community is working towards this new direction at all levels be it cooling or AC factors or cutting down power units. A mark like 20 Megawatts per 1000 petaflops is indeed a different one altogether.
Srinivas Tadigadapa, Director, Enterprise Sales, Intel – South Asia reckons energy efficiency in supercomputing endeavours as an extremely important piece. “To that extent Xeon has always been best-in-class when you compare performance and efficiency in the same zone. Intel is leading with efforts in this area and at this balance.” Incidentally, Intel has been seen to provide the processors for the largest share (82.4 percent) of TOP500 systems.
TSUBAME-KFC, which is a heterogeneous supercomputing system developed at the Tokyo Institute of Technology (TITech) in Japan, topped the Green-most supercomputer the list with an efficiency of 4.5 gigaflops/watt. Each computational node within TSUBAME-KFC consisted of two Intel Ivy Bridge processors and four NVIDIA Kepler GPUs. In fact, all systems as observed in the top ten of the Green500 used a similar architecture, i.e., Intel CPUs combined with NVIDIA GPUs.
Not so incidentally, the latest June 2014 Green500 list indicated a marked inclination for heterogeneous supercomputing systems at the top, with the top 17 spots of the list gobbled by heterogeneous computing systems, securing six more spots at the top over the previous list. This 55 per cent increase says a lot about use of supercomputing systems that deploy computational building blocks that consist of two or more types of computing brains or a mix of traditional processors (CPUs), graphics processing units (GPUs), and co-processors.
It’s hard to ignore the ‘Tango’ effect when you see that for heterogeneous systems the average energy efficiency was around 1,938 MFLOPS/watt, but it dragged only up to 743.32 MFLOPS/watt for the measured homogeneous systems.
Ground Clearance vs. Glass Ceilings
Now, Piz Daint, the greenest petaflop supercomputer in the world as per the earlier count, submitted the highest-quality Level 3 measurement (was recognized consuming a total of 2.33 MW and delivering 2.7 Gflops/W) while Tianhe-2, the fastest supercomputer in the world according to the Top500 list, achieved an efficiency of 1.9 gigaflops/watt. Notably, it could have been a first time of sorts where the average of the measured power consumed by the systems on the Green500 nosedived in comparison to previous edition of the list.
Green measures and directions are the only way the field can move towards exascale (computing at the rate of 1000 TFLOPs or higher) without using an unrealistic amount of power, if we can interpret Prof. Bondhugula’s prognosis here.
As a press note from the Green500 team also pointed out – A decrease in the average measured power in tandem with an overall increase in performance can only be seen as encouraging step on the way to exascale.
There was also an assumption that if TSUBAME-KFC’s energy efficiency can be maintained for an exaflop system, the extrapolation to an exaflop supercomputer has been seen to hit below 300 megawatts (MW), in an unprecedented way.
Of course, the 222-MW power envelope was assessed in all candour as a long way away from DARPA’s target of an exaflop system in the 20-MW power envelope, as per the report.
For now, one can find silvery silicon linings in the idea that the avowed purpose of the Green500 is to provide a ranking of the most energy-efficient supercomputers in the world. This initiative asserts well that for decades, the notion of “performance” has been synonymous with “speed” (calculated in FLOPS,) and this particular focus has led to the emergence of supercomputers that rake in egregious amounts of electrical power and produce so much heat that extravagant cooling facilities must be constructed to ensure proper operation. That also turns out resulting in TCO hikes with other metrics like reliability, availability, and usability conveniently tucked away in oversight or oblivion.
TSUBAME-KFC in the latest green ranking, stayed at the top and remaining the only system above the 4,000 MFLOPS/watt mark. In the last November ranking, it hovered at around 27.78 KW on total power criteria with a 4,503.17 MFLOP/watt. This time the figures move to 4,389.82 MFLOPS/watt at total power of 34.58 KW.
The 10 greenest machines from the previous Green500 have meanwhile, remained in the top 15 of this Green500 list, and were joined by similarly heterogeneous architectures.
What is hard to shrug is that Tianhe-2, the fastest supercomputer in the world according to the Top500 list, also managed to squeeze inside the 50 most energy-efficient supercomputers in the world list at number 49, possibly due to its heterogeneous architecture, accelerated by Intel Xeon Phi, and at an efficiency of 1,902 MFLOPS/watt despite its size.
Tianhe-2, a supercomputer developed by China’s National University of Defense Technology, is the world’s No. 1 system with a performance of 33.86 Pflop/s (quadrillions of calculations per second) on the Linpack benchmark, according to the 43rdedition of the twice-yearly TOP500 list of the world’s most powerful supercomputers.
In short, there is room and hope for bridging the gap between the two supercomputing races.
Green500 is one remarkable effort to eke out awareness to other performance metrics like performance per watt and energy efficiency for improved reliability but the world needs more stakeholders putting the word ‘green’ in the radar (or blinkers of) of supercomputing race.
Everyone and anyone (from drivers to passengers to hitch-hikers) can do his or her bit here in abridging a long-running chasm.
This is perhaps where closing a split will invoke more awe. Shrinking this gap between two forward-bound vehicles after all, would be worth some giga/peta goose bumps. Or something that we call- Epic.