Constructing the ASCI computational grid

Abstract

The Architecture descriptionThe ASCI grid architecture provides the software infrastructure for an integrated information and simulation environment for the nuclear weapons complex in 2004. The Distributed Resource Management (DRM) project provides grid services to higher level applications. The initial implementation has two development thrusts. Fkst, ASCI requires a core integrated environment for job submission to classified ASCI platforms at the three weapons laboratories:Los Alamos (LANL), Lawrence Livermore (LLNL), and Sandia (SNL). G1obus provides these core services, with some extensions for specific ASCI requirements. Second, a software layer above the core services provides a set of common capabilities that support higher-level problem solving environmen( PSES). These capabilities include complex work management and resource brokering. This architecture demonstrates the following features: The use of Globus services for accessing resources in an independent ASCI environment Kerberos-based authentication, grid information authorization and accesss control A Globus interface to the two-tier architecture of Sandia's Computational Plant (Cplantm), a distributed computing resource built of commodity parts CORBA-based work management services to support the coordinated use of multiple resources and complex task sequencing CORBA-based criteria-driven resource brokering that extends job requestiresource matching to support software resources and load-balancing Java-based client for generic work requests Web-based monitoring Integration of the grid infrastructure with CORBA, Java, and Globus-based PSES and tools.The DRM grid services address three aspects of the shift from platform-centric to network-centric computing: software resources, work flow, and co-scheduling. Software resources serve a large class of users and problems where many resources are suitable, and the primary interest is how soon the results can be obtained. Some simulation codes are run by a number of users at different sites, and must be available for many of the grid computing resources. Multiple versions of these codes are maintained to support different users and to ensure reproducibility of results. Requests to "run code X" are satisfied by software resources and brokering. Workflow services manage complex task sequencing in the network environment, analogous to the complex scripts often submitted in a platform environment. Workflow can coordinate computational processing steps, computation and visualization, computation and data movement, or other resource usage. Subtasks are scheduled independently.'Sartdiit is a multiprogram laboratory operated by .%ndia Corporation, a

What makes workflows work in an opportunistic environment?

“…Nimrod-G [23] is a cost and deadline based resource management and scheduling system. The Accelerated Strategic Computing Initiative Grid [24] distributed resource manager includes a desktop submission tool, a workflow manager and a resource broker. In the ASCI Grid software components are registered so that the user can ask 'run code X' and the system finds out an appropriate resource to run the code.…”

Optimizing the ASC WAN: evaluating network performance tools for comparing transport protocols.

“…The Advanced Simulation Computing Wide Area Network (ASC WAN, formerly known as the ASCI WAN) was developed to connect supercomputers within the US Department of Energy National Laboratories, allowing the various labs to share resources [1]. The locations which utilize this network are Sandia National Laboratories, Albuquerque, NM (SNL); Sandia National Laboratories, Livermore, CA (SNL-CA); Lawrence Livermore National Laboratories, Livermore, CA (LLNL); and Los Alamos National Laboratories, Los Alamos, NM (LANL).…”

Globus and planetlab resource management solutions compared

“…At deployment, depending on available resources and planned applications, specific service implementations can be chosen and deployed, often in conjunction with other GT-based components. GT is in production use across VOs integrating resources from 20-50 sites [9][10][11][12][13] with thousands of computational and data resources, and is expected to scale to 100s of sites with 1000s of sites as a future goal.…”

“…Grid applications are often compute-intensive [10][11][12], although some also consume significant amounts of disk and/or network bandwidth as a result of focusing on, for example, integration of large-scale data repositories (data grids [30][31][32][33], virtual observatory [34]), collaboration [35], or the control of scientific instruments ( [36,37]). PlanetLab services are generally network-intensive and rarely have significant CPU demands.…”

“…In some environments, the broker may forward the job under its own identity or under the identity of the VO with which the user is associated. Existing functionality: Numerous VO-level schedulers (or 'meta-schedulers'), some domain-or application-specific, have been developed by various groups (e.g., Nimrod-G [60], GrADS [61], DAGman and Condor-G [62], ASCI DRM [11,63], EU DataGrid). In addition, GT includes a co-allocation broker, DUROC [64].…”