[LCRC Accounts] Yearly Allocation Request from CouplingMechOpt
Hello, A yearly allocation for the LCRC cluster has been requested with the following updated information: Submitter/PI: Jay Larson Project Name: CouplingMechOpt Division: MCS Project title: Mechanics and Optimization of Coupling in Multiphysics and Mulitscale Systems Associated funding: DOE Other Systems: ALCF Science: The objectives of this project are to explore Study (and implement) mechanisms generally useful Study, prototype, and (with hope!) implement in production infrastructure useful to coupled systems, including: * Schemes for intercomponent load balance, including load diagnostics, and mechanisms for addressing intercomponent load imbalances through malleability of component cohorts * Studies detailing intercomponent data dependencies in parallel coupled and performance measurement and modeling to inform load balance, component cohort sizing, and * Simulators of large coupled systems to forecast coupled model throughput Project description: Coupled systems--both multiscale and multiphysics--are increasingly relevant to modeling for basic scientific research and for highly applied engineering applications. The advent of ultrascale platforms comprising up to million-way parallelism begs the question: What kind of applications are amenable to this monumental computational capability, and how might we use this unprecedented computational power efficiently? Multiphysics and multiscale models comprising multiple subsystems in mutual interaction are a likely application for these platforms. The challenge then becomes developing infrastructure for managing what are likely to be time-evolving component loads while minimizing processor idle time in components awaiting coupling input from other components. As a first step towards a solution for intercomponent load management for parallel coupled models, we are developing a mechanism for multicomponent load management that includes support for dynamical resizing of component cohorts at runtime. The framework is built to complement and extend the Model Coupling Toolkit (MCT; http://mcs.anl.gov/mct), key coupling middleware used most notably in the Community Climate System Model. The resource requested is for prototyping of the aforementioned intercomponent load management system, and extensive sanity and performance testing of the system in a set of use cases ranging from simple “toy” problems up to (assuming all goes well) a coupled climate model such as CCSM. MCT is highly portable and routinely compiles/runs on fusion. MCT uses MPI and any Fortran compiler available on fusion. We expect we may need only minimal system consulting support to perform the work outlined. We are requesting access to fusion so that we have a medium-sized cluster to test out ideas and code. Typical job sizes would require small numbers of processors (<=64) for periods of 15 minutes or less; larger jobs would run for correspondingly shorter periods. Thanks to LCRC support, we have had early success prototyping and performance testing of a load-balancing system that enhances MCT to support malleable model coupling; that is, coupling between parallel models whose member components vary in the number of MPI processes used during execution. This work was published in the ICCS 2011 conference proceedings (see reference at end). At present, we are working on a refined testbed application and more sophisticated predictive models to improve the intercomponent dynamic load balance mechanism, work that will continue in FY2012. At present, the testbed application still uses statistical models that summarize timing data of the Community Climate System Model. The current allocation should be sufficient to cover this ongoing portion of our work. We are also creating a more computationally complex testbed, which will embody solver algorithms as proxies for earth system model components. We almost certainly will require greater resources to experiment on this more realistic testbed system. Reference: [1] Kim, D., Larson, J.W., and Chiu, K. (2011): "Toward Malleable Model Coupling," Proceedings, ICCS 2011, Published as Procedia, 4, 312-321. Project URL: http://mcs.anl.gov/mct Current FY Hours Used: undetermined amount New FY Requested allocation: 120000 Q1: 30000 Q2: 30000 Q3: 30000 Q4: 30000 Justification: Thank You, The LCRC Accounts System
participants (1)
-
accounts@lcrc.anl.gov