Supercomputer boffins warm to clouds

A straw poll conducted at the SC10 supercomputing conference indicates that big HPC shops may be eyeing cloudy setups.

HPC folks have had to share their servers, storage, and networks for many decades because of the high cost of building large-scale machines, but in general they've been loath to share and have therefore been skeptical about using cloudy infrastructure as the foundation of their systems.

But that may be changing.

Platform Computing, one of the pioneers of grid clustering software for supercomputing, did a straw poll among attendees at the SC10 supercomputing conference in New Orleans last November. The company asked 100 IT managers at research labs, government facilities, educational institutions, and in various industry sectors, what they thought about using public, private, and hybrid clouds to run HPC workloads. Of those polled, nearly two-thirds said they were evaluating using cloudy infrastructure, and of those who were not, the vast majority said that they would at least consider the idea in the next year.

Among the IT execs at HPC shops that have monkeyed around with cloudy infrastructure, 74 per cent said they had a positive experience using cloudy infrastructure, and 45 per cent said they had a positive experience when cloudbursting their existing workloads from their own systems out to HPC utilities.

It's not clear how many told Platform Computing they had so-so or bad experiences – but this is IT and this is people, so there is no way that everyone could be happy.

HPC shops are considering moving at least some of their workloads to private, public, or hybrid clouds (which mix the two) for a number of different reasons. About a third of those polled who were testing or using clouds in production said they were interested in HPC clouds because such setups give them a more-flexible infrastructure, while 15 per cent cited the ability to provision resources more quickly. Twelve per cent said it let them solve problems they could not afford to tackle in the past, and 9 per cent said it was more cost-effective to use cloudy HPC than trying to build and maintain clusters themselves.

Of those who are using cloudy HPC, 15 per cent are bursting existing workloads onto public clouds, 36 per cent said they were building internal HPC clouds for better sharing infrastructure, and 23 per cent said they were hosting applications on public clouds right from the start.

In August 2009, Platform Computing did a similar straw poll of 103 IT executives at the International Supercomputing Conference in Dresden, Germany, and those surveyed were not as interested in using clouds to support HPC jobs. In that older survey, 28 per cent of the IT managers polled said they expected to install a private cloud within the year.

And in a survey of 95 executives from the SC09 supercomputing event in November 2009, the results of which were made available last February, 82 per cent of those polled did not think they would be bursting workloads out to public clouds, but 45 per cent said they were considering setting up an internal cloud to see how the technology might be applied to HPC work.

As it turns out, many more actually did something, according to the latest Platform Computing straw poll.

Public-cloud juggernaut Amazon Web Services, a subsidiary of the eponymous online retailing giant, believes that HPC shops will eventually come around to using public clouds to do their number crunching and run their simulations. But in the short run, Amazon also thinks that history will repeat itself.

EC2 got its start among developers who were frustrated by the time and money it took to deploy a physical server in their own IT organizations, and the myriad scientists and quants who want access to iron on a short-term basis will no doubt find it easier on their nerves – and possibly their budgets – to use public clouds to do the number-crunching for some of their workloads.

To that end, in July 2010 Amazon tweaked its EC2 compute cloud to run in parallel mode with 10 Gigabit Ethernet networking and dedicated hardware . In that same month, Peer 1 Hosting, a traditional and cloudy server-slice provider, fluffed up a CPU-GPU hybrid cloud aimed at HPC shops so they could test out how well GPUs might accelerate their simulations. In November 2010 Amazon followed suit, adding GPU coprocessors to EC2 slices.

Even supercomputer labs themselves, under pressure, are getting into the act. Back in December, Daresbury Laboratory in the United Kingdom teamed up with HPC system integrator OCF to peddle time on Daresbury's CPU-GPU hybrid cluster to outside buyers. And Microsoft has been encouraging HPC customers to run their code on its Azure cloud in recent months, as well. ®