HPC APPLICATIONS: DESKSIDE HPC
➤
cluster to a single workstation using a Tesla GPGPU card, Gupta says, the researcher saw processing times reduce from 6.7 to 2.6 hours. ‘GPUs don’t accelerate everything, but they do accelerate a lot of scientifi c computing and engineering applications,’ says Gupta, adding that the types of application the GPUs can handle has moved beyond vector-based tasks: ‘Our new GPUs have L1 and L2 caches, allowing unstructured programmes to fi t on the GPU. In computational dynamics, for example, most of the problems being solved today are unstructured grid problems, rather than classical vector problems, but they still map well to the GPU.’ Gupta believes that the affordable computing power offered by GPUs has changed HPC for the better: ‘GPUs democratise supercomputing, because everyone can have one, in the same way that everyone can have a workstation, but not everyone can have a cluster. Even large IT companies are limited in terms of how much computing they can put in a data centre, but this approach allows them to get more bang for their buck.’
Integrators such as SGI and Cray see the value of offering different types of hardware to different customers, as Cray’s Bolding explains: ‘If the person using the simulation is a developer, they’re going to need to try out their code on different platforms, and so it’s important to provide them with fl exibility in terms of system confi guration. We provide a standard X86 Intel confi guration, as well as Nvidia accelerator options – which are harder to program for, but are potentially very powerful – and a range of other choices. You can’t do HPC at the deskside and only have one offering.’ Despite Nvidia’s Cuda framework, GPUs and parallel architecture in general remains diffi cult to code for, but Gupta expects more and more applications developers to program in parallel as competition between them drives progression. ‘Many customers are running codes developed in-house, with universities in particular developing their own applications,’ he says.
In GPU-based deskside HPC, as is the case for the personal clusters offered by SGI and Cray, the focus is on allowing users to retain control of their applications: ‘There’s a lot of work going on in the life sciences, including biochemistry, bioinformatics,
34
and traditional gene sequence analysis applications,’ explains Gupta. ‘So-called gene banks are becoming better and better at collecting genetic information and the databases behind them are getting really large. Now, to do any kind of search or similarity matching can take a very long time. Traditionally, such sequence analysis used to run on a desktop, but [due to the volumes involved] many researchers have been running them in data centres for a few years. The GPU has enabled these researchers to move the tasks back to the desktop,’ he says, adding that fl exibility and control over the resources used is important to many users, particularly academics.
Horses for courses SGI’s Tanasescu explains the challenges the company faced in developing its Octane III deskside HPC system: ‘The challenges of developing the system were mostly in the packaging of it,’ he says. ‘Power, cooing, and noise were the areas we put effort into. We used regular motherboards and nothing too specifi c in terms of hardware, but the value-add was in terms of packaging and taking care of cooling, all within a noise envelope suitable for an offi ce system.’
‘Everyone said the mainframe was dead when the client server was introduced in the Seventies. Still today, people buy mainframes’
For the next generation of systems, SGI is working on packing more compute power into the device by putting more cores and more nodes into the same power envelope. ‘We’re also working on hybrid clustering, which is combining Linux and Windows,’ explains Tanasescu, adding that this offers users increased versatility from their systems. ‘Most desktop applications run on Windows, whereas most simulation applications prefer Linux. Currently, we need to run either entirely in Windows, or entirely in the Linux environment, but a hybrid workstation cluster would allow the ability to move from one environment to the other and back again, all within a single workfl ow.’
SCIENTIFIC COMPUTING WORLD OCTOBER/NOVEMBER 2010
The third area of SGI’s research efforts focuses on creating application specifi c confi gurations, incorporating SSDs, accelerator cards, or energy-effi cient nodes as required to meet the demands of a very specifi c task: ‘You can, for example, reduce the number of nodes required by enhancing the I/O capacity of each through the use of SSDs,’ he explains. ‘A strong I/O system can be used for applications such as data mining and data warehousing, all contained in one box. When we develop these appliances, we do so with a specifi c application in mind. It’s a development onwards from the personal cluster.’ In the near future, small and medium-sized companies will be able to purchase a cost- effective, compact system, tailored to the precise needs of application – an appliance rather than a supercomputer.
The future...?
The success of deskside HPC stems from the fact that true HPC iron is expensive, and only really necessary for the most demanding applications: ‘The reality, in particular for manufacturing, engineering, and the life-sciences, is that the applications they work with don’t scale too high,’ says Tanasescu, listing CAE, computational chemistry, bioinformatics, reservoir simulation, and seismic programs as examples of areas that might fall into this market. ‘They don’t need a petafl op system, but they scale in the range of 32 to 64 cores or maybe fewer. For such applications, a system like Octane III is the right platform. For teams of one to fi ve, it can even be used as a workgroup cluster.’ But what of the future? While cloud offerings are not suitable for many customers at the moment, will the market for deskside HPC shrink as cloud computing becomes cheaper and more trustworthy? Tanasescu doesn’t think so: ‘You know, everyone said the mainframe was dead when the client server was introduced in the Seventies. Still today, people buy mainframes.’ Cloud HPC, he says, will not replace deskside HPC within the growing mid-range market. ‘The percentage will change, and more people will use the cloud, but technical engineering companies will probably use both, with desktop HPC for everyday use and cloud for heavy jobs.’
www.scientifi
c-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 |
Page 49 |
Page 50 |
Page 51 |
Page 52 |
Page 53 |
Page 54 |
Page 55 |
Page 56