Wow, what an exciting start for my SC10 trip to New Orleans. This morning I joined over 250 HP customers, partners, and staff who arrived at SC10 a few days early to attend the HP-CAST (HP Consortium for Advanced Scientific and Technical Computing) user group meeting. This was the 15th HP-CAST meeting and broke all previous attendance records. Those of you who are Twitter fans may have seen some of my tweets @marchamilton (http://twitter.com/marchamilton) during the day from the event. Most presentations will be posted on the HP-CAST web site, which can be reached from HP’s HPC landing page. Here is a summary of day 1.
HP’s theme of accelerating academic, research, and commercial customer’s HPC initiatives came thru loud and clear. While some sites seem intent at simply throwing as many GPUs as necessary into a cluster to buy their way into a Top 10 spot on the Top500 list by brute force, some HP customers have spent a lot more time and effort preparing for what is likely an inevitable shift to heterogeneous computing. Professor Satoshi Matsuoka of Tokyo Tech described his four years of work with accelerators and GPUs, starting with TSUBAME1.0 in 2006 to today’s TSUBAME2.0 system. Achieving more than 1PF sustained performance, over 30x that of the original TSUBAME1.0 system, while maintaining the same floor space, power, and cooling envelope, TSUBAME2.0 sets new records for GPU cluster efficiency.
Lest you think GPUs are just a fad, Tom Read of Nvidia showed a world map with the locations of the 1000+ GPU clusters currently installed worldwide. One such HP system, even newer than TSUBAME2.0, is the Georgia Tech Keeneland system just installed at ORNL. Based on the same HP SL390s servers used in TSUBAME2.0, the Keeneland cluster, while not as large, uses even more advanced GPU technology. Keeneland is one of the largest clusters today using the Nvidia M2070 GPUs which just recently started shipping and offering 6 GB of GPU memory, double the 3 GB of TSUBAME2.0’s M2050 GPUs. Witness to the completeness and value of HP’s factory integration, Keeneland completed a Top500 class full system Linpack test less than a week after being delivered. Clusters of this size have often taken many weeks or months of post-delivery setup and integration before being able to complete a system-wide Linpack benchmark.
Based on all the major CPU & GPU vendor’s roadmaps, it is hard to argue against the growing adoption of heterogeneous computing. However, Nvidia, AMD, and Intel are all taking different approaches which was very clear in today’s roadmap presentations by each of these vendors. AMD, taking advantage of their x86 processors, is placing GPUs on-die with their Fusion architecture and delivering increasing levels of on-chip x86 core and GPU core integration over the coming years. Not having the software infrastructure head-start of Nvidia’s CUDA environment, AMD is placing a much bigger bet on the open source Open-CL environment for programming its GPUs. Intel, with its Intel MIC co-processor, is taking yet another approach to co-processing. The Knights family of Intel MIC chips promises to deliver a much more x86-like development environment to programmers, although it is available today only in a very limited volume software development kit with the first full production MIC chips still some ways off. The GPU/co-processor space is clearly still in its infancy and promises to be a hotbed of innovation as we move from Petascale to Exascale computing over the rest of the decade.
Another area of HPC that is undergoing rapid change is HPC storage. From constant advances in the use of SSD-Flash technologies to the rapidly changing Lustre ecosystem, HPC storage capabilities continue to grow. While HP’s 3PAR acquisition received perhaps the most storage PR this year, HP’s Sean Cochrane updated the audience on how HP has incorporated the IBRIX technology acquired last year into HP’s X9000 HPC storage offerings. Where X9000, with its large global namespace and segmented file system excels, is HPC environments with many to many (compute node to file) or one to many environments needing extremely high file create and random I/O rates.
HP-CAST attendees also heard details about many future HPC technologies being worked on in HP-Labs. From the memristor to HP Labs <a href="http://www.hpl.hp.com/research/intelligent_infrastructure/">Exascale technologies, HP Labs is working on a host of new technologies that are likely to be delivered in future HP HPC systems.
HP-CAST continues on Saturday with parallel in-depth tutorial tracks from both HP staffers as well as from many of our partners. Since you will have plenty of time during SC10 to enjoy New Orleans in the vendor events, if you are in town Saturday, consider stopping by the JW Marriott on Canal St., catching one of the HP-CAST tutorials, and meeting some of HP’s HPC team. We’ve had a focused HPC hiring effort across all functions, from HPC sales to marketing to engineering, and we are not done yet!
That’s a wrap for today.