“Early operations mode” describes the status of two NSF-funded systems that are on track to support a wider range of user than is traditionally served by elite-level supercomputing. Wrangler is the Texas Advanced Computing Center (TACC) system that we reported on last week, so now we turn our attention to Comet, the petascale supercomputer readying for launch at San Diego Supercomputer Center (SDSC).
Comet is the outcome of a $12.6 million grant from the National Science Foundation (NSF) to field a system that expands access and capacity across traditional and non-traditional research domains and accommodates the long-tail of science, a concept that refers to more modest-scale jobs that make up a significant portion of research. This move towards broader engagement speaks to NSF’s larger cyberinfrastructure strategy too, a topic we’ll return to after a brief rundown on Comet.
The Dell-integrated cluster occupies 27 racks, with 72 nodes per rack for a total of 1,944 compute nodes. Each node is outfitted with two Intel Xeon E5-2600 v3 12-core processors (running at 2.5GHz), 128 gigabytes of traditional DRAM and 320 gigabytes of local flash memory. A total of 46,656 cores contribute to a peak performance of 2 petaflops.
To optimize capacity for modest-scale jobs, each rack has a full bisection InfiniBand FDR interconnect from Mellanox, with a 4:1 over-subscription across the racks. Comet also claims 7.6 petabytes of Lustre-based high-performance storage, plus 6 petabytes of durable storage for data reliability, as well as 100 Gbps connectivity to Internet2 and ESNet.
The standard Xeon nodes will provide the bulk of the compute capability, but Comet also has 36 GPU nodes, equipped with four NVIDIA GPUs and two Intel processors. And soon it will also have large-memory nodes, outfitted with four Intel processors and 1.5 TB of memory. The heterogeneous configuration will enable Comet to more optimally target specific workloads, such as visualization, molecular dynamics simulations or de novo genome assembly.
Like SDSC’s Gordon supercomputer, as well as TACC’s Wrangler, Comet will become part of the XSEDE (eXtreme Science and Engineering Discovery Environment) system. Comet replaces Trestles, which entered production in early 2011 under an earlier NSF grant.
One of Comet’s more interesting features is its support for high-performance Single Root I/O Virtualization (SR-IOV) at the multi-node cluster level. Comet’s use of SR-IOV will allow virtual sub-clusters to run applications over InfiniBand at near-native speeds. This ‘secret sauce’ lowers the entry barrier for a wide range of researchers by permitting them to use their own software environment, but still attain supercomputer-level performance.
“Comet is really all about providing high-performance computing to a much larger research community – what we call ‘HPC for the 99 percent’ – and serving as a gateway to discovery,” said SDSC Director Michael Norman, the project’s principal investigator. “Comet has been specifically configured to meet the needs of researchers in domains that have not traditionally relied on supercomputers to solve their problems.”
Both Wrangler (at TACC) and Comet (at SDSC) were funded by NSF’s Track 2 program, which formed in 2006 with the mission to award $30 million on a competitive basis every year to deploy a new supercomputer into XSEDE. (Former SDSC User Services Consultant Glenn Lockwood provides a helpful summary of these now archived awards.)
Currently the NSF is investigating a new funding methodology in keeping with its vision for Advanced Computing Infrastructure. As part of the Cyberinfrastructure Framework for 21st Century Science and Engineering (CIF21), the program focuses “specifically on ensuring that the science and engineering community has ready access to the advanced computational and data-driven capabilities required to tackle the most complex problems and issues facing today’s scientific and educational communities.”
In the most recent solicitation for HPC system acquisition (posted Feb. 14, 2014), the NSF called for “new and creative approaches to delivering innovative computational resources to an increasingly diverse community and portfolio of scientific research and education.”
The shift toward “a more inclusive computing environment” is further clarified in the program guidelines with some of the more salient paragraphs copied below:
Recent developments in computational science have begun to focus on complex, dynamic and diverse workflows, which integrate computation into all areas of the scientific process. Some of these involve applications that are extremely data intensive and may not be dominated by floating point operation speed. While a number of the earlier acquisitions have addressed a subset of these issues, the previous solicitation NSF 13-528 and the current solicitation emphasize these aspects even further.
…Consistent with the Advanced Computing Infrastructure: Vision and Strategic Plan (February 2012), the current solicitation is focused on expanding the use of high-end resources to a much larger and more diverse community. To quote from that strategic plan, the goal is to “… position and support the entire spectrum of NSF-funded communities … and to promote a more comprehensive and balanced portfolio …. to support multidisciplinary computational and data-enabled science and engineering that in turn supports the entire scientific, engineering and educational community.” Thus, while continuing to provide essential and needed resources to the more traditional users of HPC, this solicitation expands the horizon to include research communities that are not users of traditional HPC systems, but who would benefit from advanced computational capabilities at the national level. Building, testing, and deploying these resources within the collaborative ecosystem that encompasses national, regional and campus resources continues to remain a high priority for NSF and one of increasing importance to the science and engineering community.
The results of this solicitation were unveiled in November with the announcement of “Bridges,” focused on problems related to data movement, at the Pittsburgh Supercomputing Center and “Jetstream,” a cloud-based system, co-located at the Indiana University Pervasive Technology Institute and the Texas Advanced Computing Center. The new resources, valued at $16 million, are anticipated to come online in early 2016.