towards exascale
‘for us it’s part of our technology roadmap to provide a very scalable solution up to 40 Gigabit using Ethernet as a standard interface.’ According to Ford, very low latency is the
key to scalability: ‘Te wider the network, the more hops you are going to take to get across it. Today, Gnodal’s offering is the fastest, the lowest latency switch in the world. Te inter- switch hop is only 66 nanoseconds – one sixth of our closest competition. Tat’s the building block you need to build these larger networks.’ But he warned that in larger networks,
other effects of Ethernet kick in; the most important of which is congestion. Here, he believes, Gnodal has a radical solution: ‘Our approach is to avoid congestion in the first place by offering a multipath Ethernet fabric.’ Activity tends to be ‘bursty’, he remarked, and at times the load will exceed the mechanisms used to orchestrate traffic across the network. Conventional techniques oſten involve a higher layer protocol, so ‘to manage the data that is already congesting, you are having to overload it with even more data to try to help the congestion’ he says. Gnodal does not use higher layer protocols
at all. Te path/load-balancing across the fabric is determined by the ASIC switches themselves. One further characteristic makes the Gnodal solution interesting to the Mont- Blanc project – it is topology invariant and not bound by traditional hierarchical networks.
Forge: a 153 teraflop supercomputer combining CPUs and GPUs, located at the University of Illinois’ National Center for Supercomputing Applications
expertise will come in. She explains: ‘For Cray, we do not build the processor node, but the networks and the infrastructure, so we will have to figure out how to build more powerful networks.’ Te challenges for Cray on the hardware side, then, are building a powerful enough network to handle traffic from these supernodes, being able to cool them, and keeping the density pretty high. But more parallelism at the node level
PEOPLE SHOULD NOT LOSE SIGHT OF THE FACT THAT EXASCALE REALLY IS DIFFERENT
Within Mont-Blanc, the interconnect may require Ethernet on one side and something else on the other. Te issues of system balance and of faster
interconnects are also very much in the mind of Peg Williams, Cray’s senior vice president of High-Performance Computing Systems: ‘In supercomputing, it is all about the balance of the system. You have got to balance the processing power, with the amount of memory, with the amount of network bandwidth that you have. If you get out of balance, one of those will become the gate and slow the system down.’ In her view, the need for energy-efficiency
will drive new processor architectures with changes in the parallelism at the node level: ‘What kind of processors do we need that can operate at reasonable power and high flop count?’ Each node is going to be pumping out a lot more traffic to the network than nodes today, so if the system is to be balanced, these more powerful nodes will require more powerful networks and this is where Cray’s
26 SCIENTIFIC COMPUTING WORLD
‘exposes the second problem: for the code developers to find and expose the parallelism in their code and algorithms,’ she says. Tis in turn will drive work in new programming models. Most current programming models – of putting one parallel process per compute core – just will not scale at this level. ‘Te problem blows up when you get a lot of parallelism at the node.’ Te memory model on the node may change as well, and that could also drive changes in the code. In this, she sees a further challenge for the company: ‘We provide programming environments for our systems. So we have to have compilers and tools that map to the new nodes. We have the soſtware problem that we have to address as well – in the sense of having the right tools for the programmers to address these new nodes.’ Alex Ramirez believes that the particular
architecture of the Mont-Blanc project will make the soſtware challenge acute: ‘Each one of the processors is going to give us a lot less performance than current processors, so we are going to have to use a lot more of them.
Tat translates into additional effort for the programmer because now the application has to scale to a much larger number of parallel threads than it had to do before. Te difference is that to achieve a performance level you have to write your application to a thousand processors, not a hundred processors.’ Among the ways in which soſtware is being
addressed on the Mont-Blanc platform, is for the French company Allinea to supply the debugger. Jacques Philouze, of Allinea, says: ‘ARM is moving from the low end and they want to address the high end, and that is a big challenge for them – to go to an architecture with double floating point. Change of architecture is a nice challenge for us and we are pleased to work with Mont-Blanc for the duration of the project.’ He points out that Allinea is already supporting ARM whose chips are finding application in the financial services market. At present the contract is commercial in confidence but he expects there to be public announcements soon showing the financial markets are investing in the ARM architecture. For Pete Beckman director of the Exascale
Technology and Computing Institute at the US Argonne National Laboratory, the interests of the end users need to be taken into account in devising the architectures of Exascale systems. ‘Inside the research programme, three “co-design” centres have been launched: on combustion; materials; and one in nuclear energy reactor design,’ he says. Te idea is that these centres will look at the algorithms needed for the science that they want to achieve and what that would mean for the architecture of the computer systems. ‘As we look at Exascale, what features can we modify or change that
www.scientific-computing.com
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 |
Page 45 |
Page 46 |
Page 47 |
Page 48