June 24, 2014

IBM Offers Glimpse into Future of HPC Investments

Nicole Hemsoth

This week at the International Supercomputing Conference (ISC14) one of the more frequent topics of conversation has been the revised role that some expect IBM might play following the Lenovo x86 business acquisition—and how their investment in high performance computing might change following the transition.

There’s little wonder this topic is on everyone’s lips, especially when reminded of IBM’s steady role on the TOP500 list, which was just announced. Big Blue still enjoys a sizable share of the Linpack benchmark’s finest with its 35.2% share of the overall list (just a tick below HP). But of course, these are systems that will be going the way of the dinosaur as IBM refreshes its HPC strategy in the wake of the many changes around OpenPower investments, the Lenovo deal, and heightening competition from developments at the chip and accelerator level. While this list and HPC is still important to IBM, they’re seeing this segment as something that dwarfs the Top500–in fact, it stands as the antithesis to the FLOPs-centric view.

IBM has been vocal this week during presentations and individual meetings about the steadfast nature of its commitment to supercomputing, and finally, we’re starting to get a sense of what this means beyond mere words. It appears that over the course of the year, we’ll see that lip service to ongoing HPC investment play out in the form of some major announcements that tie together innovations in OpenPower and a fresh approach to the “workflow” concept as the dominant theory guiding future system design.

What’s most compelling is that we can expect this vision, which incorporates the activity around OpenPower (and naturally, the Power architecture roadmap), a move away from a compute-centric emphasis, and a focus on real-world commercial HPC operations to materialize in the form of actual product announcements in 2014 and a more narrowly defined product and strategy roadmap for 2015.

At the heart of this is will be that HPC workflow concept, which carries a few central tenants. First, the management, acquisition and manipulation data cannot be removed from the HPC conversation—the big data phenomenon is too powerful to ignore and it factors prominently in nearly all of the actual end user HPC applications IBM’s technical computing team has evaluated—even as far back as their planning phases for the Sequoia system, says Turek. The problem, he explains, is that system designers are still rooted in the outmoded cluster approaches that ignores the fact that compute is the simple part—it’s the data that represents the challenge.

With further heterogeneity in systems, increases in programmatic complexity, more bottlenecks fed by memory and bandwidth, and now, sheer data overload, Turek says the only possibility is to take a workflow approach to designing future architectures—one that is data-centric and moves away from the “disparate components” present of system design.

Workflow, as IBM’s Dave Turek told us in a detailed conversation around the questions at ISC, is a term that is defining their HPC system strategy going forward since it includes the increasing data-centric requirements the technical computing team at IBM is finding dominate the FLOPS-defined needs at centers in government and industry. While he says that academic centers have slightly different algorithmic needs that might tie in better with world of Linpack performance, the dominant need for other HPC centers is around dramatic increases in integer performance—not floating point.

“If you look at the operational folks HPC, their focus is so often on servers. They get their interconnect, storage, and other elements from a collection of vendors and cobble these things together. But why would anyone do that, why would anyone decompose the offerings and center it all on servers when so much of the real computational activity is married to the network, to storage, to the manipulation of data?” With its coming announcements later in the year, IBM expects to bring everything together into the right package that exploits the latest advancements in memory, storage, networks, and of course tie it to the right server package in a programming model that is “utilitarian”.

Their goal with the coming announcements later this year is to mesh all of these concerns and potential solutions and leverage the support of the OpenPower foundation. With the many memory, storage and other vendors on this list, one can speculate that they’ll be looking to their own on-package memory, integrated storage, and data locality solutions to move into a system-wide approach. We’ll learn more around supercomputing, but it’s difficult to ignore the roadmap items from Intel that just became clearer this week in terms of their use of Micron’s memory technology at the chip level and the new OmniScale fabric to push out development at the interconnect and networking layers. Turek sounded quite confident that whatever they’re cooking will have a more seamless value-add for HPC centers commercial and otherwise and at least reading between the lines, it seems that putting the data closer to the computer in a comprehensive, system-level way could be their ticket to take an enterprise product and give it an HPC gloss with some added HPC acceleration via the OpenPower possibilities.

Again, this is roadmap speculation. But there’s still the question of the present. Is IBM already losing ground because of the Lenovo questions and lack of future for its core HPC technologies, including BlueGene, that will be allowed to wither?

“Look at our Open Foundation members list,” was Turek’s answer. With the new string of announcements expected later in the year, Turek said they’ll be harvesting a number of technologies that have fed their existing HPC strategy but again, the way these will arrive is in the form of more integrated, broader, workflow-defined approaches that fit the data-centric needs of actual end users versus the more compute and FLOPS-centric form that is so often associated with high performance computing centers.

There are not many details about how the actual transition will take place since it’s not been finalized, even following today’s presentations at ISC around the topic of IBM’s future—a future that other Big Blue execs are echoing. Turek says that IBM will continue to operate as if there’s been no change, then sometime later this summer, will provide more details about the actual mechanics of the Lenovo changeover.

Turek reiterated that “IBM is centered more generally on a business orientation that focuses on innovation, but that innovation is difficult to do in a commodity space when the entrances are closed and there’s a single supplier who can dictate a lot of the terms for future development. OpenPower will buttress our general approach going forward with respect to HPC.”

Share This