towards exascale Bill Dally, chief scientist at Nvidia


as appears on a benchmark. Probably the most difficult is power – because we don’t just want to build an exascale machine, we want to construct one that can run in a machine room with a total power budget not much higher than today. We have tentatively adopted the goal of exascale at 20MW and to do that requires almost 100:1 improvement in performance per Watt, which is a very tough goal to meet. One of the main things we’re working on is finding ways of improving the energy efficiency of computing so that we can close that gap and reach exascale computing in a power envelope that will be acceptable. ‘Another main challenge is reliability as


‘T


an exascale machine is going to have many more components than is seen today. As


here are three main challenges to reaching exascale – as defined and sustained performance on real and meaningful problems, and not just


components become smaller, the same size fluctuations can cause larger disturbances. We need to develop techniques that make our machines very robust in the presence of transient errors, and build highly-reliable


PARALLEL MACHINES ARE


ALREADY HARD TO PROGRAM AND IF YOU SCALE EVERYTHING UP AND KEEP IT AS BUSINESS AS USUAL IT’S GOING TO BE INTRACTABLE


machines out of somewhat unreliable components. The third major challenge is one of programmability. Parallel machines are already hard to program and if you scale everything up and keep it as business as usual it’s going to be intractable.


Gilad Shainer, senior director of HPC and technical marketing at Mellanox


what needs to be delivered or created throughout the process. Both hardware and software must be addressed, and from our perspective, the focus is on both. Hardware and software needs to be co-designed, as without doing so, it will be very difficult to develop larger scale


‘T


systems that can be fully utilised. Saying that, we should not create proprietary tightly-coupled solutions. Still, we want to have future systems used for a broad range of applications and industries. If exascale will not go into the commercial market then the cost of an exascale system will become unaffordable. ‘There is no doubt that all the issues that will come across are


solvable. If you look back there were always limits or barriers that people would say couldn’t be exceeded, but then suddenly someone would have a breakthrough. The same will happen here, as nothing is impossible. The big question is how to build these new systems effectively. There are a couple of voices out there in the industry who believe that exascale systems will not be used for a single full-scale application, but rather be used for multiple applications running on smaller portions of the systems. Others talk about exascale being able to be used by single applications, enabling jobs to be scaled up. ‘I agree with the latter opinion as I believe that exascale


machines needs to enable single applications to utilise the entire system. We want to be able to solve bigger problems and handle the large amounts of data being collected to analyse it.’


www.scientific-computing.com FEBRUARY/MARCH 2011 27


he best way to look at the exascale challenge is to break the problem into smaller steps and then look into


‘It’s clear


to me that the solution is a heterogeneous supercomputer where the nodes incorporate CPUs to do the serial portions of the code and GPUs to do the throughput sensitive portions. I think that’s something parts of the community, and certainly the Chinese, have embraced. Realistically as GPUs are so much more energy efficient measured in Watts per Flop than their counterparts, this is the only way we can build machines of this scale with the energy efficiency we need. There are certainly those who are dragging their feet on this transition and an important message to get across is that to get to exascale we have to take the leap and embrace heterogeneous computing – it’s the way of the future.’


Page 1  |  Page 2  |  Page 3  |  Page 4  |  Page 5  |  Page 6  |  Page 7  |  Page 8  |  Page 9  |  Page 10  |  Page 11  |  Page 12  |  Page 13  |  Page 14  |  Page 15  |  Page 16  |  Page 17  |  Page 18  |  Page 19  |  Page 20  |  Page 21  |  Page 22  |  Page 23  |  Page 24  |  Page 25  |  Page 26  |  Page 27  |  Page 28  |  Page 29  |  Page 30  |  Page 31  |  Page 32  |  Page 33  |  Page 34  |  Page 35  |  Page 36  |  Page 37  |  Page 38  |  Page 39  |  Page 40  |  Page 41  |  Page 42  |  Page 43  |  Page 44