The HPC community’s preoccupation with 1000X increments is well-documented. These thousand-fold speed-ups are parlayed into flashy systems, unveiled roughly every ten years or so. IBM’s Roadrunner had barely broken the petascale barrier in 2008, when the HPC community issued a collective, “what’s next?” Enter extreme computing.
According to the Council on Competitiveness, “Extreme Computing can be defined as pushing the boundaries of what can be currently achieved with High Performance Computing (HPC) and is 500 to 1,000 times more capable than current high-end systems.”
The Council, a public policy organization comprised of CEOs, university presidents and labor leaders, believes that extreme computing “will help the U.S. sustain its competitive edge in basic sciences, engineering, and research and development – and when applied to industrial and commercial applications will fuel US innovation and economic growth.”
To explore the implications of the emerging post-petaflop era, the Council on Competitiveness was awarded a $914,000 grant from the Department of Energy, the US government agency now driving the nation’s exascale roadmap. The new endeavor complements the work that the Council has done with small to medium-sized enterprises, under its National Digital Engineering and Manufacturing Consortium project (NDEMC, pronounced “endemic”).
At SC12 in Salt Lake City, the same week the DOE funding was announced, HPCwire sat down with Council on Competitiveness Senior Vice President Cynthia R. McIntyre to discuss the merits of the project. Dr. McIntyre has been leading the Council’s High Performance Computing (HPC) Initiative since 2008.
According to her, the project is well aligned with their mission to encourage public-private partnerships to advance innovation in the US. They believe this is an effective way to leverage federal investments – in computing and software development – to achieve scientific return that can be transformed into economic gains on the industry side.
The DOE grant came about through meetings of a sub-group of the HPC Advisory Committee. They saw that certain areas of computational technology were not developing at a rate that supported innovation and acceleration. For example, they noticed a lack of new application software development. Software developed in the public sector is available, but it is not being captured by industry. The codes coming out of research could be more widely deployed if they were made industry-hardened, however.
The group also discussed the need for software codes that could be employed to make petascale systems feasible and useful, and they wanted to know if their current codes would scale to work on petascale systems. Concern was voiced over whether the ISV community would support emerging platforms. In McIntyre’s words, “there is a lead lag situation.”
The companies that were involved in these talks saw an opportunity for the feds to lead and help advance the application software for commercial use. They talked about having more flexible access to the highest-end computing platforms. Currently they can gain access through collaborative research agreements, CRADAs, which allow companies to partner with national labs on research problems that can be jointly addressed and have mutual benefit. The mechanism is in place, but it can be a time-consuming affair. Naturally, companies are interested in streamlining the process; they’d like to establish a faster connection to the federal labs.
These large companies know that DOE will be moving toward exascale, but in order to plan for the future and create their own roadmaps, they need a better idea of what’s going on at the federal and research level. This will help them begin to understand the kinds of problems that they can solve on next-generation systems and how their codes will make the adjustment. Basically, they want to have a conversation with the DOE, to be briefed on what’s being talked about and what’s being decided.
“The industry is interested in exascale,” says McIntyre, “but they need an opportunity to hear about what’s going on, to reflect on that and to then to begin to think strategically about when this platform comes online or the technologies become available how’s that going to impact what they do in terms of their computing.”
Because of its connection to key stakeholders – CEOs, university heads and labor leaders – and its engagement with national labs, the Council is in a position to convene these parties to hear about what the roadmap is for exascale coming out of the DOE.
“The OEMs need a much more organized and focused engagement to understand how the technology, and the plans for the technology, are evolving,” explains McIntyre. “We would like to bring the OEMs to the table with the DOE to talk about the kinds of projects or problems that could be addressed with such a system if certain elements are in place. This is what we will be doing with the DOE grant – talk about the issues of software development, codes that the DOE will have – and find out is there a mechanism for commercializing those codes for industrial use.”
This grant covers a period of three years during which the Council will publish the conversations and findings in a series of reports and case studies. While it’s too early to speculate on the specifics of the schedule at this time, the DOE has submitted a report that should provide further details once it is made public.
The path to exascale presents numerous difficult challenges, especially in regard to the limitations of current architectural designs. There are memory walls, power walls, and issues of fault tolerance. When scaling current systems no longer suffices as a way forward, progress hinges on breakthroughs. But this requires something that has been in rather limited supply: the political will to support funding that could enable such breakthroughs.
This week the DOE pushed back its exascale timeframe projections by two to four years, which means the first US systems aren’t expected to come online until at least 2020, but more likely, not till 2022. And that’s assuming the agency’s 2014 budget proposals are accepted by Congress.
In aiming to create a better understanding of the “drivers and benefits of Extreme Computing – in particular as it relates to our country’s industrial and commercial competitiveness in the global economy,” one has to wonder if the strength of the Council’s conclusions will help convince US funding agencies to loosen their purse strings. On the other hand, do we really need more proof of the merits of a well-funded national HPC strategy?