As a runner-up to the much larger Supercomputing Conference held in November, the International Supercomputing Conference (ISC) in Dresden, Germany is a convenient platform for delivering mid-year HPC product announcements and company news. There was plenty to go around this year. For those keeping score, the event is also the summer venue for the semi-annual shuffling of the Top500 list.
And let’s face it. ISC is also a good excuse for the HPC crowd to get together, drink a little German beer and unwind. Check out the picture of the Venusbrass-Band woman with come-hither looks promoting Tuesday’s Exhibition Opening Party at the top of the ISC social events web page. My caption would have been “Girls Who Love Guys Who Love Big Iron.”
But I digress. My true intention here is to summarize the more noteworthy announcements at ISC and add a few thoughts of my own. Here we go.
One of the biggest announcements at ISC was IBM’s introduction of the new Blue Gene/P, the company’s next-generation supercomputing technology. The new platform will take the reigns from Blue Gene/L and carry the Blue Gene architecture into petascale territory. In our feature article this week, we dig a little deeper into Blue Gene/P and describe how IBM took Blue Gene to the next level.
While IBM is a leader in the supercomputer realm, it really would like leadership in all price bands. The company is especially interested in the high-growth cluster segment, which encompasses department, division, and workgroup systems. To stay competitive here, you have to learn how to weave commodity components and software stacks into a product that stands out from the crowd. The IBM Cluster 1350 product line is the platform aimed at these higher growth segments.
IBM’s approach has been to partner with ISVs and work through the value added reseller channels to get more direct access to customers. The ISV partnerships have become especially valuable to IBM to get pre-engineered solutions targeted to specific HPC domains like CAE, financial services and life sciences.
“If you want to be a leader in the middle and lower price bands, you have to have a good channel,” said Herb Schultz, Deep Computing Marketing Manager. “IBM is not going have its traditional sales force knocking on doors to get a $50,000 cluster sale. We have to look at channel opportunities.”
IBM’s recently announced Power6 microprocessor also has a play in the HPC portfolio. The new processor is currently one of the most powerful CPUs on the market (interesting factoid: each core in the Power6 outperforms the Deep Blue chess-playing computer of 1997) and has the advantage of integrated accelerators for certain types of compute-intensive workloads — the decimal floating point accelerator for business applications and the Altivec math accelerator for 3D modeling. The new p570 server, which sports the Power6 chip, can accommodate a variety of mid-market HPC customers.
A new Power6-based HPC server is also in the works for the middle of 2008. It will represent the Power5 upgrade to the p575 and is to be used as a building block for large, highly computationally-dense clusters. This platform will act as a bridge to the Power7 technology. That processor is the basis for the PERCS architecture being developed for DARPA’s HPCS program. According to IBM, PERCS platforms are slated to achieve 10 petaflops in the 2010 timeframe.
At that point, IBM will have two petascale architectures — Blue Gene and PERCS — that target somewhat different application profiles. The Blue Gene design is based on large numbers of low-power nodes with specialized interconnects, and is suited for workloads that exhibit fine-grained parallelism. PERCS is based on larger SMPs with more memory, which makes it more suitable for problems that need more single-threaded performance. PERCS will also encompass a variety of software innovations — the P in PERCS stands for productive. The goal is to develop a software system that overcomes barriers in petascale programming and systems management.
IBM wasn’t the only vendor talking peta in Dresden. Sun announced its new petascale supercomputing architecture there as well. The company’s Sun Constellation System is a product line that uses an ultra-dense 3456-port InfiniBand switch to connect large numbers of Sun Blade 6000 servers. The Sun switch, which contains Mellanox silicon, is the key technology that provides the interconnect performance necessary for a petaflop-scale cluster. It uses a fat tree network topology to reduce the number of switches and optimize communication latencies between nodes. The servers themselves come in a three flavors: Opteron, Xeon or Sparc T1 processors.
Compared to IBM’s Blue Gene/P, the Sun design follows the conventional cluster model more closely, using standard chips (with the exception of the T1) and InfiniBand technology. A lot hinges on the success of the InfiniBand uber-switch to pull it all together. If the company is able to achieve supercomputing performance at prices that undercut Cray XT4s and the IBM Blue Genes, Sun may be able to find an opening in that elusive capability supercomputing market for commercial HPC users.
The Texas Advanced Computing Center (TACC) will be the first customer to install a Sun Constellation System. The TACC machine, named Ranger, is spec’ed at 500 teraflops (peak), and is expected to be installed by the end of the year. Sun is hoping the machine can be built by November, in time to make a run at the number one spot on the next Top500. There’s some suspense around the schedule, since it’s not assured that the quad-core Opterons used to build Ranger will be available in that timeframe.
SGI introduced a new high-end cluster platform at ISC — the Altix ICE (Integrated Compute Environment). The product is essentially a scaled out version of SGI’s Xeon-based Altix XE platform. The Altix ICE is Xeon-based as well, but is bladed and uses RAS technology from the company’s Itanium-based Altix platforms. For example, it uses the power supply and water chilled doors found in the Altix 4700. The ICE nodes are lashed together with 20 Gbps InfiniBand. SGI wanted to give a single-system image kind of feel to the platform, so they incorporated a hierarchical management network that hides some of the clustery-ness from the end-user. SGI is targeting scale-out applications that typically run on big clusters, large SMP systems, or even Cray Red Storm type systems.
An single Altix ICE rack will hold up to 512 quad-core processors, providing over 5 teraflops per rack. “With a couple of these racks you have a good shot at getting on the Top500 list,” said Bill Mannel, senior director of SGI Server Marketing. According to him, the platform can scale out to 40-60 racks and the price-performance becomes better as you add capacity. He said that on a list price basis, they’re cheaper than everybody except Dell. Mannel estimates that most Altix ICE systems will end up in the $250,000-plus price range.
Intel Corp. delivered a couple of newsworthy items in Dresden. Surprisingly, neither one was about microprocessors.
The first was a product announcement of Intel Connects Cables, an optical cable that is meant to compete with traditional 24-gauge copper cables used for InfiniBand interconnects in HPC clusters. Optical fiber cables don’t have the weight, size, and bending radius problems associated with big copper cables, so offer a lot more manageability in the datacenter. And signal quality is much better with optical media, especially at distances beyond 10 meters. The distance limitations become worse as data rates rise, which is why copper is becoming more problematic for InfiniBand as systems move to 20 Gbps.
This gives optical technology a real advantage as cluster densities become unmanageable from a space and cooling point of view. Optical interconnects can enable cluster infrastructure to spread out, allowing for more flexible datacenter setups. According to Intel, the new cables can span distances up to 100 meters without loss of signal strength.
“This distance issue is a big deal,” explained Tom Willis, general manager of Intel Connects Cables. “Right now with 8 to10 meter copper cabling, people are telling us they can only connect about 1000 CPUs at the new 20 gigabit per second double data rates. Going up to 100 meters, you’ll be able to support thousands of CPUs — on multiple floors.”
Willis said their optical cables become cost-competitive with copper beyond about 20 meters. (Glass fiber itself is cheap; the expensive components are the optical transceivers at each end.) The new cables were demonstrated at ISC in partnership with other HPC vendors, including IBM, HP, Dell, Sun, Voltaire, Mellanox, SGI, Cisco, Bull, and Microsoft.
The other Intel news was the introduction of their Cluster Ready program. Intel Cluster Ready defines a system specification that acts as a contract between hardware vendors and software developers. Using the spec, OEMs and system integrators can certify their clusters and ISVs can formally register their software. At time of deployment, integrators, ISVs or buyers can run an Intel Cluster Checker program to verify that the system is compliant.
“What the customer really wants is a cluster with the characteristics of an enterprise server,” said Dr. Herbert Cornelius, director of the Intel Advanced Computing Center. “This is what we are targeting with Intel Cluster Ready.”
No details of the specification have been released*. So at this point it’s difficult to gauge its impact on the cluster landscape. But the idea of defining reference standards for clusters would seem to be a good approach to help move these architectures into more software-friendly platforms. For what it’s worth, SGI and Dell are the only cluster manufacturers currently signed up for the new program, although many of the major HPC software vendors — ANSYS, MSC.Software, LSTC, CD-adapco and The Mathworks — are already on board.
Even though Gates & Company offered no product announcements at ISC, Microsoft technical fellow Burton Smith keynoted the event. His talk focused on the importance of bringing parallel computing into the mainstream IT community. We previewed his keynote address in an interview last month.
Also of note — two clusters running Microsoft’s Windows Compute Cluster Server (CCS) made it to the Top500 list this year. A 448-node IBM BladeCenter cluster deployed at Mitsubishi UFJ Securities made it to number 193; and a 256-node Dell PowerEdge cluster installed at Microsoft’s Tukwila, Washington datacenter came in at number 106.
Keep in mind that the Top500 is not where Microsoft intends to make its living. The more traditional targets for Windows CCS are the smaller workgroup and departmental clusters. In particular, as some HPC market segments, such as oil & gas, replace desktop workstations with deskside (personal) supercomputers, CCS becomes a natural fit since it doesn’t require the level of system administration that a typical Linux-based platform does. Microsoft is continuing to position Windows CCS as a clustering-for-the-masses technology and taking advantage of the changing perception of HPC as commodity tool.
“Since [Windows CCS] became generally available in the summer of 2006, I think there’s been a shift in the definition of HPC,” said Jeff Wierer, senior product manager on the Windows HPC team. “Customers are definitely still looking for high performance, but now there’s this idea of high productivity coming into play as well. People are looking for solutions that are familiar and easy to use. Our sweet spot is where performance meets productivity.”
According to the Top500 organizers, the list had the biggest turnover in its history, as a lot of older, slower machines got booted off. The Top500 is certainly becoming a more exclusive club. Entry to the list required a 4-teraflop system this time around; just a year ago it was 2.7 teraflops. Total accumulated capacity has increased to nearly 5 petaflops, which is nearly twice the 2.8 petaflops at this time last year. IBM and HP systems still dominate the list — IBM in total capacity and HP in number of installations.
A lot of the Top500 sites are public sector organizations. But quite a few commercial entities are listed as well, some referenced anonymously to protect their identities. For example, Geoscience company E has an IBM BladeCenter ranked at number 70; Semiconductor company P has eight HP clusters at 125 through 132; and Financial Services company N has three BladeCenter clusters at 297 through 299.
Even a Government Classified entry is listed with two HP clusters at numbers 139 and 140. But it seems unusual that no classified systems are listed higher than that. I’m guessing that at least a few 10-teraflop-plus covert supercomputers are out there. In fact, I wouldn’t be surprised if one of those machines will be used to perform data mining on this very article… umm, not that I have anything to hide.
* Actually, Intel has released the Intel Cluster Ready specification. It is located at http://softwarecommunity.intel.com/UserFiles/en-us/File/1311/Intel_Cluster_Ready_Specification_LinuxV1.0_20070524.pdf. I’m sure I’ll be covering this initative in future issues. [07/05/07]
As always, comments about HPCwire are welcomed and encouraged. Write to me, Michael Feldman, at [email protected].