One of the more heavily attended presentations of ISC’09 was the “Cloud Computing & HPC: Synergy or Competition?” session, which took place on Wednesday morning. The interest from ISC attendees seems to reflect the industry’s current obsession with all things cloudy. There was a good balance of skepticism and optimism on the panel, but overall I came away feeling they never really addressed the session’s question.
Representatives from HP (Richard Kaufmann, CTO Scalable Computing Infrastructure Organization), Sun Microsystems (Marc Hamilton, VP HPC & Cloud Computing), Microsoft (Dan Reed, Scalable & Multicore Computing Strategist), the Jülich Supercomputing Center (Thomas Lippert, Director, Jülich Supercomputing Center), Google (Robin Williamson, Engineering Director), Amazon (Simone Brunozzi, Amazon Web Services Technology Evangelist), and Yahoo (Dr. Sanjay Radia, Senior Architect, Hadoop Project) gave their interpretation of the cloud phenomenon, and participated in a panel discussion at the end to field questions from the audience.
There was general agreement on the benefits of cloud computing: elastic capacity, pay-per-use model, platform abstraction, economies of scale, and built-in fault tolerance. Unfortunately — and maybe significantly — there didn’t seem to be much consensus about whether the clouds would usurp traditional HPC infrastructure as the platform of choice.
In particular, the reps from the traditional cloud providers — Google, Amazon and Yahoo — didn’t directly address how general-purpose clouds would evolve to address the needs of high performance computing. They did mention frameworks like MapReduce and Hadoop as being suitable for processing extremely large data sets in a highly parallel manner. In particular, Simone Brunozzi highlighted Amazon’s Elastic MapReduce Web service, which is specifically designed for data-intensive apps like data mining, machine learning, financial analysis, scientific simulation, and bioinformatics. But no one in this group delved into performance issues or the need for more specialized infrastructure geared toward HPC.
The group from the HPC contingent (HP, Sun, Microsoft, Jülich) pointed out that cluster interconnect performance, in particular, will need to be addressed before cloud computing gets much traction with supercomputing users. Amazon EC2 gets halfway there by offering multiple types of servers, called instances. The instances are set at different price points depending upon the server profile — CPU horsepower, memory capacity and I/O capability. But as of today there is no option for, say, InfiniBand-equipped servers. The general consensus from HPC practitioners is that the lack of a high-performance fabric in these general-purpose clouds will restrict adoption. “It’s all about the interconnect,” noted HP’s Richard Kaufmann.
There are also data security and privacy issues, but they apply to a range of applications, not just HPC. These concerns are well known to cloud providers and presumably will be addressed more completely as a greater number of users demand them.
Sun’s Marc Hamilton brought up the issue of public and private clouds, noting that if your organization can achieve its own economy of scale with regard to computing capacity, private clouds may be the way to go. According to him, public and private clouds can live side-by-side, but only if interoperability (i.e., cloud API standards) are developed. Even the public cloud arena would benefit from these standards, since users would rather not be locked into a single provider. Hamilton, for example, noted that while the cost of entry into the Amazon cloud is very low, the cost of exit may end up being high. In fact, he blamed the lack of interoperability in Sun’s Network.com offering as a major contributor to its demise.
Thomas Lippert took on the role of the cloud skeptic, especially in regard to the kind of cutting-edge supercomputing that goes on at places like Jülich. He believes the cloud model won’t support leadership supercomputing. And it’s not just the performance issue. The whole supercomputing ecosystem at that level is so specialized (support, hardware, and software) and the lifecycles of such systems (3 to 5 years) are so limited, that the cloud model wouldn’t apply at all.
Lippert is probably right here. Although elite supercomputing is partially based on commodity hardware and software, the resulting infrastructure and applications are highly customized. Moving the grand challenge application people to the cloud would be like trying to convince Formula One racers to take the bus. Efficiency is not the driving force here.
On the other end of the spectrum was Microsoft’s Dan Reed. He believes it’s inevitable that cloud will engulf high performance computing, or at least the vast majority of it. The driver will be economics, inasmuch as the cloud makes computing and storing data in bulk extremely inexpensive. The idea is that just as commodity components crowded out specialized HPC architectures, cloud platforms will eventually edge out traditional HPC infrastructure.
Reed thinks much of the resistance to cloud computing by HPC users is actually sociological, not technological. Outside of the supercomputing realm, most users don’t care about infrastructure. They’re being paid to focus on their applications and produce results. Most of them would like to avoid dealing with the inner workings of the platform. As Reed put it: “Successful technologies are invisible.”