A trio of keynote presentations from Intel, Google and Microsoft at the PEARC19 conference in Chicago on July 31 charted out the likely future of academic and high-performance computing in the cloud. While each company and presenter carried a distinct message about the opportunities and challenges to moving more open research to cloud services, each also held that cloud providers are learning from the HPC community and adjusting their products and models to make the transition more attractive.
PEARC19, in progress in Chicago this week (July 28-Aug. 1), explores current practice and experience in advanced research computing including modeling, simulation and data-intensive computing. The primary focus this year is on machine learning and artificial intelligence. The PEARC organization coordinates the PEARC conference series to provide a forum for discussing challenges, opportunities and solutions among the broad range of participants in the research computing community.
Redefining HPC
In her presentation “Redefining HPC,” Patricia Damkroger of Intel looked at the paradigm shift that’s moving data analytics and AI into the cloud.
“We’ve talked about HPC going to the cloud for at least a decade,” she said. “It’s still not mainstream, but I think that’s changing … The biggest driver is data.”
For varied reasons, she explained, organizations as different as CERN and the Department of Defense have found loading data into the cloud to be a useful expansion of their internal compute capacities that allows collaborative access and maintains internal security, respectively.
Data are also a central need in AI, in which training data have become massive and the infrastructure required for transparency and accuracy expand. AI and HPC, she argued, are converging—or at least ought to.
“We need … to know what the AI is doing to the data. We also need to make sure we have review boards and security built in … The other thing we really need is the inclusive part,” she said citing the problem that much medical research has not been gender or race inclusive and so the results don’t always fully represent the patient population. “AI is going to have to have that full data, or it’s not going to be accurate.” She cited San Diego Supercomputer Center’s Expanse, the Texas Advanced Computing Center’s Frontera and Pittsburgh Supercomputing Center’s (PSC’s) Bridges-2 as examples of upcoming systems that will play roles in this convergence.
Damkroger shared the podium with Nick Nystrom of PSC, who gave the audience the first public presentation of the center’s new Bridges-2 system. The NSF announced the award for Bridges-2 in June. Bridges-2, built in collaboration with HPE, will feature Intel’s 10nm Ice Lake processor along with other Intel CPUs.
“We’ve been working on this for a while,” he said. “This [system] was a convergence of HPC, AI and data.” Designed for use by “new community” researchers with little or no computing experience and employing the first instance of Intel’s Omni-Path Architecture, Bridges-2’s predecessor, Bridges-1, runs common applications that make it cloud-friendly. The system, Nystrom added, is able to run HPC modeling and simulation alongside common tools such as Jupyter as well as Spark and big-data workflows, bridging work that requires the strengths of HPC and cloud. Bridges-2 will expand on that capability.
Future Is HPC in the Cloud
Google’s Ross Thomson’s keynote “Future Is HPC in the Cloud” surveyed the company’s offerings via Google Cloud Platform to enable true HPC in the cloud.
“There’s always a place for the giant computers people use to do massive simulations” for users with $100 million to fund top-500 systems, he said. But for users—or collections of users—who don’t need such a large system, “you can get a lot of computing done for $100 million on Google Cloud.”
He cited Google Cloud’s capability to provide virtual systems configured to each user’s required size, enabling them to scale up or even scale down without losing their investment as their needs change. HPC in the cloud, he added, can accelerate discovery by reducing queue wait times for large-batch workloads as well as relieve compute-resource limitations.
Are There Closets in the Cloud?
In “Are There Closets in the Cloud?” Microsoft’s Tim Carroll charted the history of academic clusters from dozens of systems in literal closets spread across campuses to the sophisticated—and in many ways optimized—campus systems now in operation. He noted that while some 70 percent of academic HPC centers employ cloud computing, only 10% of their jobs run in the cloud.
“The idea is to get more tools in more people’s hands, so that they can do good things with them,” Carroll said. For that to happen, both HPC and cloud providers will need to make cultural changes. “One of the things [in which] I think the cloud providers have done tremendous disservice to ourselves and the community is time and cost being the only metrics that matter in this space.” In some cases, they are; but in public research, ownership over systems, dual use in computer and domain science and different funding models than in the public sector can all make that simple calculus inaccurate.
“All of these machines serve a dual purpose and are not simply a utility,” he said. “One cannot underestimate the impact of that.”
Among others, Carroll cited the National Oceanic and Atmospheric Administration (NOAA)—which employs some of the most powerful HPC systems in the world. NOAA is seeking to move its global weather forecast code and capabilities into the cloud. This allows outside collaborators and even citizen scientists open access to spur innovation.
“The tipping point was access, not price,” Carroll said. “[The] evolution and revolution is about opening up computation to domains of science that have never had access before … That’s a really important point to consider when we get a little wrapped around the axle these days about whether the cloud is right for HPC.”
Carroll recommended that HPC users carry out four activities to chart out how they can determine the cost of using the cloud. Step one is to plan, identifying and inventorying workloads that might run well in the cloud. Second, running both obvious workloads and those that may not run as well will generate real data to provide realistic performance expectations. Collaborating with cloud providers can help smooth out cultural differences and produce more accurate estimates. And finally, cost estimation should come at the end of the process rather than the beginning, because workflows drive the true cost.