On Friday IBM announced a formal collaboration with GENCI to prepare scientific codes for the path to exascale. In July IBM announced its second European POWER Acceleration and Design Center (PADC), this one in Montpellier, France. It’s also the locus for the GENCI work. Last November IBM opened the first PADC, this one in Germany (Jülich Supercomputer Center), as well as a Big Data Innovation Hub in the UK (STFC). Of course IBM Centers of Excellence have sprouted around the U.S – some supporting CORAL – and also in China.
“This is going to be a connect-the-dots exercise over the next year or so as you see us announce more centers. It’s a further indication, contrary to popular belief, that when we divested ourselves of the Intel-based business IBM never left the HPC space; we just kind of restructured a little bit,” said Dave Turek, vice president of high performance computing at IBM in an interview with HPCwire.
So far, the young effort seems to be bearing fruit. When the first PDAC opened, (Jülich Supercomputer Center), OpenPOWER had 70 members. Today it has roughly 140. Turek notes quickly the PDAC centers are run in concert with OpenPOWER Foundation cofounders NVIDIA and Mellanox and they are indeed helping expand the OpenPOWER ecosystem but involvement is hardly restricted to OpenPOWER members. Individual IBM collaborations such as with GENCI may also benefit OpenPOWER, but are focused on more generally advancing efforts to get to exascale, he said.
“Somewhere between 20 to 30 percent of all the participants in OpenPOWER now have an HPC emphasis and there are a number of activities being pursued in partnership with those companies. But there’s a lot of work OpenPOWER is doing independently of IBM to support their own HPC ambitions,” said Turek. “I would say that’s reasonably characteristic of some of the players in China who certainly are using our technology, licensing our IP, using IBM design services, in some cases we expect they will use elements of IBM software stack, but the overall solutions they are trying to put together have very particular kinds of local characterization.”
GENCI is the lead organization spearheading France’s drive towards exascale. The new IBM-GENCI agreement will focus on scientific application software. IBM expects the collaboration to extend well beyond the 18 months cited in the press release, “perhaps to seven or eight years or more” because the path to exascale is long. “IBM will have people dedicated to GENCI so it’s not just empty promises. There are real people assigned to GENCI and working on this fulltime. It’s a serious commitment on our part to leverage the kind of expertise we find at GENCI and advance the exascale cause broadly from an international perspective,” Turek said.
The vast majority of software produced by the collaboration, according to Turek, will make its way into open source. “Where there’s a commercial code that needs to be ported or optimized but that will probably be executed mostly by IBM in concert with the commercial software provider,” he added.
Talking about the ostensible exaflops goal, Turek said “We don’t actually have a flops centered view of this market any longer. When you really look at the incorporation of the analytics and frankly even some of the classical HPC applications, the presence of more flops does nothing for you. We’ve known that since at least 2012 during when we did a self examination of the portfolio of applications running on Sequoia (Blue Gene) and discovered the way to have gotten more performance would have been to do more in terms integer performance and memory bandwidth, two architecture features that are reflected in the POWER8 and future power designs.”
The IBM-GENCI collaboration will also attempt to take full advantage of OpenPOWER-based innovations such as the connection of NVIDIA GPUs accelerators to POWER processors through the high-speed NVIDIA NVLink interconnect, as well as how Mellanox EDR switches can exploit IBM’s Coherent Application Processor Interface (CAPI) to dramatically improve solution performance. Experts from GENCI and IBM will also work on understanding the evolution of programming models, considering alternatives to MPI and OpenMP application program interfaces for shared memory multiprocessing programming.
“I have been very careful as we laid out these centers – optimization centers, design centers, power-based technology centers, centers of excellence in the U.S. to support the coral program – to orchestrate the work so there is no redundancy in terms of what’s going on. This is an international collaborative effort to kind of divide and conquer,” said Turek.
For example, “we’ve made sure that no work going on at GENCI was being echoed by one of the centers of excellence in the U.S. and conversely. So the centers of excellence that are part of the CORAL project will actually derive benefit from the centers that we set up in Europe and we expect to announce more centers in the coming months to make this progressively more and more international,” said Turek.
Ensuring HPC is broadly responsive to regional requirements is critical but unfortunately rare, said Turek. Citing an HPCwire podcast (Industry Leaders on the Peril and Promise of NSCI) with panelists from Cray, HP, IBM, Intel, and SGI, he noted wryly “the attendees on your call all have American accents. We’ll count Gabriel [Broner, SGI] as one of our own even though he’s from Argentina but he’s been here way way long enough.”
There is a difference of perspective from geo to geo and there are concentrations of expertise in certain geographies that create natural opportunities for collaboration. That’s important, Turek said, because it is easy to forget, especially for the big companies operating in HPC, the kinds of impediments that people around the world have to deal with when they contemplate the deployment of this technology – simple things like continuous availability of electricity, the cost of power, widely varying regulations, even different cultural attitudes toward HPC use.
“We sometimes make pronouncements as if they are applicable to a homogeneous characterization of humanity across the planet when they are not. For example, there are different attitudes toward cloud as applied to research and LS in Europe than in the U.S.,” noted Turek.
Accommodating those differences is a practical necessity in growing HPC use and IBM’s business. “By setting up centers like this around the world and using them as a kind of instructional mechanism for our systems designers to make sure what they build actually resonates well on an international basis. As time passes you can get our a world map and start connecting the dots and just see how diverse this is going to turn out to be.”