Here’s a collection of highlights, selected totally subjectively, from this week’s HPC news stream as reported at insideHPC.com and HPCwire.
>>10 words and a link
AMD Fellow argues acceleration makes more sense than manycore;
http://insidehpc.com/2008/06/09/amd-fellow-acceleration-makes-more-sense-than-manycore/
AMD releases new quadcores, Opteron SE;
http://insidehpc.com/2008/06/11/amd-releases-new-quads/
IBM puts water cooling inside 3D stacked chips;
http://insidehpc.com/2008/06/09/water-inside-ibms-new-approach-to-cooling-stacked-chips/
NetApp revs storage products with HPC in mind;
http://insidehpc.com/2008/06/10/netapp-aims-at-technical-computing-with-new-producs/
Appro’s 40 Gbps InfiniBand cluster;
http://insidehpc.com/2008/06/10/appro-shows-off-40-gbps-ib-cluster/
New multicore info hub lights up on the Web;
http://insidehpc.com/2008/06/08/new-multicore-info-hub-online/
SGI to offer Moab suite;
http://insidehpc.com/2008/06/11/sgi-to-offer-moab-suite/
Canada’s HPCVL installs new SPARC T5140;
http://insidehpc.com/2008/06/10/hpcvl-selects-sun-sparc/
NetEffect 10GbE supports MS NetworkDirect;
http://insidehpc.com/2008/06/11/neteffect-10gbe-supports-ms-networkdirect/
Roadrunner uses Panasas storage;
http://insidehpc.com/2008/06/11/panasas-and-roadrunner/
Ireland and Turkey join PRACE;
http://insidehpc.com/2008/06/11/prace-consortium-grows/
>>IBM’s Roadrunner Breaks Petaflop Barrier
IBM announced Monday morning that they have officially broken one petaflop with their latest installment of the Roadrunner supercomputer. The new machine, destined for Los Alamos National Lab later this summer, performed the Linpack benchmark at 1.026 PFLOPS [1.026 quadrillion floating point operations per second]. Once released, it will edge out its BlueGene/L cousin on the TOP500.
What does it take to reach 1-PF? IBM used 13,824 AMD Opteron cores coupled with 12,960 Cell enhanced double precision processors. All this at a cost to LANL of ~$120 million.
The most interesting aspect of this story is the hybrid nature of the Roadrunner machine. We’ve all seen musings in the HPC industry surrounding hybrid platforms. Few have been brave enough to make a run at constructing a usable hybrid platform. Several hybrid platforms exist in specific markets, e.g., genomic science and financial services. None, however, were designed specifically for general workloads. Is this a sign of things to come? It’s too early to tell. However, it certainly makes sense to build a hybrid platform and subsequently enable users to place operations on the most advantageous hardware.
For more info on the Roadrunner 1-PF news, read HPCWire’s full writeup. There was also a nice piece in the New York Times that we discussed at InsideHPC.com.
>>AMD signs deals to help coders use parallel gear
News this week from The Register about a pair of deals that AMD has made with parallel tools providers aimed at helping accelerate the effectiveness of developers trying to make use of new multicore chips. From the article:
First is a pact with Rogue Wave Software, maker of products that let developers run their single-threaded apps concurrently on multiple cores without needing significant code tinkering.
The terms of the deal were not disclosed, but AMD said it would use Rogue’s Hydra software suite to expand its parallel computing strategy. The companies appear to be targeting the financial services industry specifically with this one — although that may have something to do with announcing the agreement at a securities and financial markets conference in New York.
…The chipmaker also got RapidMind’s version of the multi-threaded application voo-doo to work with AMD FireStream 9170 CPUs and ATI Radeon HD 3870 GPUs.
The news isn’t on AMD’s site (yet), but is at RapidMind. The potential value of deals like these seem to be the creation of a de facto interface to multiple types of parallel hardware…kind of like an MPI, but abstracting architecture, not message passing semantics. Not that I’m suggesting these deals will create this, but the more we focus on putting technology into a fixed (small) set of tools for programmers, the better.
>>The Little Green Lizard that Could
Novell has just released some interesting deployment statistics for the HPC industry. Apparently, Novell SUSE lives behind the consoles of 40 percent of the top 50 supercomputers (via the TOP500 list). They are also happy to point out that the top three machines wear the badge of the green lizard. BlueGene/L at LLNL, BlueGene/P at Juelich Research Center and SGI’s NMCAC ICE 8200 machine all run SUSE.
Stepping back from the marketing numbers, ask yourself who has standardized on Novell’s SUSE Linux Enterprise Server platform? Namely Cray, SGI and IBM. Granted, SGI and IBM will most certainly sell subsets of their compute platforms with RedHat. This notwithstanding, the three top supercomputing vendors in the industry have essentially standardized on the same distribution. Why is this? (I’m really asking, because I don’t know.) SGI’s senior vice president of software, Irene Qualters, had this to say:
“At SGI, our focus is on high-performance computing and robust scalability, and SUSE Linux Enterprise Server is the operating system of choice for many of our Altix and Altix XE customers. As the requirements for high-performance computing continue to grow more complex across industries, the collaboration between SGI and Novell ensures that SUSE Linux Enterprise will continue to be the leading operating system for high-performance clusters that meet those new business needs today.”
Of course, this is not to say that RedHat doesn’t have its own market share of HPC. Several of the most popular cluster management stacks solely support RedHat-based distributions (usually due to their dependence on Anaconda). The Rocks folks have certainly carved out an incredible niche in the cluster-centric HPC industry.
—–
John West is part of the team that summarizes the headlines in HPC news every day at insideHPC.com. You can contact him at [email protected].