NCAR CCSM with taskgeometry support

Independent components for modeling ocean, atmosphere, land and sea-ice behavior in changing climate scenarios are coupled together in NCAR’s Community Climate System Model (CCSM). At the implementation level, each component is a separate binary resulting in a Multiple Process, Multiple Data (MPMD)...

Full description

Bibliographic Details
Main Authors: George Carr, Ian Lumb, B. Mcmillan
Other Authors: The Pennsylvania State University CiteSeerX Archives
Format: Text
Language:English
Published: 2005
Subjects:
Online Access:http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.500.2835
http://www.spscicomp.org/ScicomP11/Presentations/User/page-paper.pdf
Description
Summary:Independent components for modeling ocean, atmosphere, land and sea-ice behavior in changing climate scenarios are coupled together in NCAR’s Community Climate System Model (CCSM). At the implementation level, each component is a separate binary resulting in a Multiple Process, Multiple Data (MPMD) application. At the implementation level, some components support a hybrid mode use of the Message Passing Interface (MPI) and OpenMP. System support for applications that utilize both shared and distributed memory under separate program control is a requirement for effective use of available computational resources. In the Fall of 2000 IBM added a 'task geometry ' capability to their LoadLeveler (LoadLeveler for AIX 5L V3.2: Using and Administering) batch subsystem manager to support CCSM and other similar applications at NCAR on an AIX-based, SP-architecture system. Recently NCAR and Platform Computing Inc. have also collaborated to enhance Platform LSF HPC for AIX with task geometry support. By addressing the locality of related tasks, Platform LSF HPC for AIX manages CCSM simulations on NCAR’s Bluedawn and Thunder systems. Recently NCAR acquired an IBM Linux cluster with low-latency, high-bandwidth message passing enabled via the Myricom Myrinet interconnect. This system utilizes LSF for batch subsystem management and also supports CCSM's hybrid on Lightning's thinner nodes. After reviewing the enhancement, usage examples will illustrate how NCAR researchers use this capability to manage asymmetric assignment of CCSM components under LoadLeveler and LSF. A throughput assessment comparing LoadLeveler and LSF is