This project seeks to develop and investigate a multi-technology environment that is hybrid in its networking, storage, operating systems, programming and computing platforms, and seeks to address issues of interoperability so as to create a productive environment for scientific and visual computing. The overall project objective is to integrate heterogeneous information systems into a workable solution that may usefully be replicated at other IBM science and technology sites and to potentially identify new commercial opportunities for the comprising IBM hardware and software. It is anticipated that the resulting composite of public domain software, commercial software, commodity hardware, and high-end, specialized hardware would become a leading center for the investigation of hybrid technologies to solve a wide array of computational science applications, which would serve to highlight IBM's diverse product technologies working seamlessly and effectively and strengthen IBM's perceived commitment to friendly engagement with open standards and public domain systems. To this end, UNM researchers, in collaboration with researchers at Sandia National Laboratory, Los Alamos National Laboratory and IBM Research, plan to develop and port a number of computational applications to this computing environment and to investigate real scientific problems in areas which include the biophysics of protein folding, the quantum mechanics of materials engineering, the dynamics of atom-ion collisions, the physics of multiphase flows, and the data mining and optimization of workloads in high performance computing. The UNM team further seeks to investigate the problem of concurrent, distributed visualization of data intensive applications using IBM hardware, including the Scalable Graphics Engine(SGE), an IBM Research prototype technology. Results from this investigation should prove invaluable to other IBM high technology customers, especially those concerned with nuclear energy and defense. UNM also plans to connect this composite technology to its existing Roadrunner Linux Supercluster -- a shared resource on the National Computational Grid -- to demonstrate complete interoperability in Next Generation Internet applications such as portal-based supercomputing, remote graphics delivery and rendering, and Internet multicasting for scientific collaboration. Using software developed for the SGE and hardware/software currently used by UNM for connectivity to the National Computational Science Alliance (NCSA) Access Grid, high resolution graphical data would be shared in real-time with other scientific institutions throughout the country. In short, the proposed hybrid environment would function to prototype an integrated "spectrum of services", ranging from relatively few local clients to potentially numerous web-connected systems that would highlight the use of clusters supporting and working in concert with the IBM SP and its attached graphics and storage systems. The proposed system would evolve in stages as determined by product availability and requested funding, (See diagrams below.)
PHASE 1: This phase would involve the construction of a 32 processor Gigabit Ethernet Linux Cluster using PCI based adapters. This phase would seek to study processor and network performance, interoperability with the Roadrunner Supercluster via an SP Switch Router, and would initiate a comparison study of distributed visualization using software rendering vs hardware accelerated rendering.
PHASE 2: This phase commences with the installation of an SP Winterhawk-II system and attachment of SGEs to the Trailblazer switch at both MHPCC and AHPCC. At this point, issues such as workload characterization, scheduling, remote access, hybrid programming, and parallel I/O performance would be investigated. Scientific applications would be ported to exploit the mixed-mode parallelism and the capability for doing large-scale computation and concurrent, immersive visualization to steer and track using both the SGE and the clusters.
PHASE 3: Finally, the entire system would be migrated, as much as possible, toward complete MPI interoperability through switch-attached interconnection and the high performance gateway node. Additional issues such as remote visualization, web-based access and application performance, would round out this technically challenging proposal.