Heterogeneous supercomputing is looking more and more like the next big thing in the high performance computing world. Now that IBM has thrown its hat into the ring with its hybrid Opteron-Cell Roadrunner system, it's hard to deny that heterogeneous computing is getting some serious respect. Will HPC turn away from homogeneous architectures and go hetero?
IBM represents only the latest vendor to jump on the hetero bandwagon. Cray, Sun Microsystems, SGI and Linux Networx have all dabbled in the hetero HPC world to one degree or another. Cray is perhaps the most homo(geneous)-phobic of all the vendors. It has staked its future on heterogeneity with its Adaptive Computing vision, which imagines systems built with a combination of many types of processing engines — conventional microprocessors, vector processors, multithreaded processors and FPGAs. Sun Microsystems, in collaboration with ClearSpeed, recently deployed a 50 teraflop Opteron-CSX600 system for the Tokyo Technical Institute. The CSX600 contains a MultiThreaded Array Processor with 96 floating point engines. And finally, a variety of FPGA-flavored systems have sprung up in recent years; SGI, Linux Networx, Cray and a few other vendors offer FPGA add-ons with some of their HPC configurations.
Why is this happening? In the never-ending quest for more computational power, many in the industry already see the end in site for conventional multi-processor, multi-core architectures. After a while, just adding more processors to a system will have no effect. If a system has more cores than you have application threads, all the extra CPUs just become Lilliputian space heaters.
The heterogenous approach offers greater efficiency by using specialized processing engines that can be matched more closely with different types of application code. A specialized chip, such as a GPU, an FPGA or a vector processor, can replace 100 conventional processors for certain types of codes. So the upside potential is enormous.
But the transition from homogeneous to heterogeneous processing is likely to be a lot tougher than going from single- to multi-core. For one thing, the single-to-multi transition for conventional processors was a fairly simple process — literally, just adding more of the same. But mixing different types of processing architectures into a unified system confronts the system architect with much bigger challenges. Some of the most pressing are:
- How tightly do you couple the various processing engines — on-chip, on the board or in the cluster? One could even envision specialized processors distributed across a LAN/WAN. (Conceptually, this model already exists as Grid computing.)
- What mix of processing engines do you use? There are a lot to choose from today and I suspect more are on the way.
- What will be the ratio of the different types of processors in a system?
And then there's the central problem of software. As difficult as it was (and is) to scale applications across more homogeneous processors, it will be significantly more complex to slice up applications across a heterogeneous architecture. Heterogeneous-aware software (compilers, run-times, process/job schedulers, etc.) that intelligently maps the application code onto the available processor resources will be required for any sort of productive use of such systems. But how do we design such software? This was the question most recently posed by the director of the Center for Scalable Application Development Software, Ken Kennedy: “How do you build software tools that are scalable from a system with a single homogeneous processor to a high-end computing platform with tens, or even hundreds, of thousands of heterogeneous processors?”
With the newly announced Roadrunner system, IBM has made some initial choices about what their first hetero system will look like. Based on Opteron and Cell boards, connected with an InfiniBand fabric, IBM believes Roadrunner will be well-suited for the workloads at Los Alamos. The final 2008 deployment of the system will use a double precision version of the Cell BE; although the first deployment, starting this fall, will use the current single precision Cell. The initial focus of the project will be on the development of the all-important programming model for the hybrid system.
I recently spoke with David Turek, vice president of Deep Computing at IBM, who filled me in on what the Roadrunner architecture means to IBM. He offered his perspective on how Roadrunner relates to the rest of the company's overall supercomputing strategy, including its Blue Gene technology. Turek also expressed his opinion on Cray's own heterogeneous supercomputing model — its Adaptive Computing vision. Read the entire interview in this week's issue. (Note: we've included Cray's response to Turek as well.)
It will be interesting to see Roadrunner in action as it deploys over the next two years. A high-profile system such as this will be sure to attract a lot of scrutiny from industry-watchers. In fact, the unhatched Roadrunner has already drawn some criticism. The High-End Crusader is pessimistic about the architecture's prospects for success. Says he:
“The essence of Cell is to be a decoupled access-execute architecture with extremely coarse granularity because of the DMA. IBM has had little success hitherto in providing system software to make the Cell in isolation programmable. Adding a layer of heterogeneity by coupling Opteron and Cell fairly guarantees that Roadrunner will be a nonprogrammable machine.
“One day, history may record that Roadrunner was a minor pebble in the broad stream of heterogeneous processing.
“Of course, hetero is still our only hope.”
Wal-Mart Does HPC … Who knew?
While the high end of HPC is busy evolving to the next stage, many companies are just starting to incorporate mainstream HPC into their business model. Last week's HPC User Conference, sponsored by the Council on Competitiveness provided an interesting glimpse into the use of high performance computing for supply chain management. At the conference, commercial powerhouses Wal-Mart, Proctor & Gamble, Pratt & Whitney and Clopay Plastic Products described their HPC experiences. Wal-Mart, in particular, relies on high performance computing to help manage their daily store operations. Contributing editor Steve Conway describes the conference's panel discussion in “HPC and Supply Chain Management” in this week's issue.
Another HPC User Conference attendee, Mike Andrescavage, chief software architect of Andrescavage Software Inc., offered a few other observations on the proceedings:
— Everyone agrees that HPC has to be made simpler.
— How can we make HPC work better?
— We need better interfaces.
— Increased U.S. productivity is needed to compete in global markets.
— A new kind of 21th century infrastructure in needed.
— We need to converge on a backbone for collaboration environment.
— We have a lack of available talent for HPC.
“I'm sorry that the above only identifies short-comings. I'll leave it to the PR folks to paint a rosy future,” says Andrescavage. “The nation's premier HPC centers are being tasked to provide expertise and capacity to willing industry enterprises. It is only a minuscule percentage that embrace and profit from HPC technology.
“Fortunately, there is no rush or on-slaught of potential partners clamoring at the gates of the NSF's nine HPC centers. It seems that interfacing, and software available for each site, is different. This is definitely a major road-block. In addition, when private industry recognizes HPC's importance, will the resources, both capacity and expertise, be available.”
Andrescavage was also impressed by the Wal-Mart HPC success story: “I enjoyed Nancy Stewart's (Wal-Mart CTO), description of one of their uses for HPC — the daily analysis and recommendations for shelf and store layout. Most retail would benefit from this kind of analysis, but lack the compute power and IT expertise.”
A recent HPCwire piece, “A Critique of RDMA,” written by Myricom's Patrick Geoffray, has provoked a rebuttal from Renato Recio, Chief Engineer, IBM eSystem Networks. Recio maintains that RDMA “threatens legacy network adapters that have not evolved to its capabilities.” Recio's rebuttal, “A Tutorial of the RDMA Model, A Response to 'A Critique of RDMA,' “ may be found in this week's issue. It contains a number of graphics, so make sure you use a decent Web browser.
The HPCwire RDMA critique piece also drew a response on the Interconnects blog site. The site is maintained by Rick Merritt, Editor at Large, EE Times. Read Rick's blog piece on the RDMA critique and a reader's comment at http://interconnects.blogspot.com/2006/08/ragging-on-rdma.html.
As always, comments about HPCwire are welcomed and encouraged. Write to me, Michael Feldman, at [email protected].