http://www.cgd.ucar.edu/cseg/timing_ccsm4/README.html 16 Mar 2008
  CCSM Timing Table
  Includes Load Balance and Performance Information
[an error occurred while processing this directive] Up One Level | CSEG Home | CCSM Home  

README

The table provides a summary of the information in CCSM timing files. The raw timing files are generally available via a link in the CS/Mode column. Lots of detail information is available in the raw timing file like run length, a copy of env_conf, and other things.

Generally, "best" load balanced configurations are presented in the table. Please feel free to forward timing files for cases that aren't currently available and we'll put those up on the site. If you find configurations that perform better (in cost, throughput, or both) than what's on the table, please send us those results as well.

The README is divided in two further sections. The first section provides a summary of the table. The second section provides a discussion of output found in the raw timing table.

Back to the table.


Table README:

A summary of the table columns follows:

  • Machine - Name of the hardware that ran the case
  • Resolution - Grid that the case used
  • CSet/Mode - Component Set and any information about the mode. B is all active and A is all data. See the CCSM Users Guide for more information about the component sets. See also the env_conf at the bottom of the raw timing file for more detailed information.
  • Total PEs - total number of pes used for the run. This is the sum of the atm, lnd, ocn, ice, and cpl pes.
  • Cost - the cost in pe-hrs for this case to run one simulated year. Lower cost is better.
  • Thru-Put - the number of years of simulation that could be produced if the model ran continuously for one day.
  • component pes - Five columns, one for each component, summarizing the number of pes used. The information there is total pes, mpi tasks x omp threads, root pe : pe stride.
  • CCSM Version - CCSM version tag used for the run.
  • Date - date the run was executed in yyyy.mm.dd.

    Back to the table.


    Raw Timing File README:

    A section of the README for the Timing Tool is provided next.

    CCSM TIMING PROFILE

    The first section summarizes the case analyzed. This information is extracted from the environment and the case setup at the time the timing tool is invoked. Most of the output in this section is straightforward.

  • "lid" is the timestamp associated with the ccsm run and is part of the filename for the ccsm log files.
  • "date" is the timestamp at the time the timing tool was run.
  • A summary of the components and their pes is provided. Generally, the pes set in CCSM are expressed as the number of MPI Tasks by the number of OpenMP Threads (MPI x OpenMP). The total is the product of these two numbers.

    ANALYSIS

    This section provides an overview of the CCSM timing and contains a summary of several critical regions in CCSM. The source of these numbers is described below in the details section. Users are encouraged to review the load balance chart to verify the comparisons provided below are applicable. Letters (ie. E+F) reference sections in the load balance chart.

    1. Overall Metrics

  • Model Throughput is the number of years of simulation that could be produced if the model ran continuously for one day.
  • Model Cost is the cost in pe-hrs for this case to run one simulated year. Lower cost is better.

    DETAILS

    CCSM is a complex system. Four tightly coupled physical components are running sequentially and/or concurrently. They communicate via the fifth component, the coupler (or driver). The communication pattern is largely set by science requirements.

    env_conf

    Appended at the end of each timing file is a copy of the case env_conf file. This generally provides addition insights into various modes or other physics that might have been turned on during the run. These can have a large bearing on the timing of a run.

    Back to the table.