Direkt zum Inhalt | Direkt zur Navigation

Personal tools

Sektionen
You are here: Home » Computing » M4HR

M4HR

last modified Sep 07, 2015 02:11 PM
The Multi-Model Multi-Member High-Resolution (M4HR) experiment

In workpackage JRA1 of the IS-ENES2 project a set of coordinated multi-model, multi-member (M4) coupled cimate simultations is developed and run on high resolution (H4). The purpose of this experiment is to address the computational performance of coupled Earth System Models in a systematic manner by definition, collection, and dissemination of metrics that allows for a robust assessment of the real model performance and intercomparison across different models and HPC systems.

Participating models and further software components

The following five ESMs have been identified for participation in the M4HR experiments:

ESM nameInstitutionComponents
ARPEGE-NEMO Météo-France / CERFACS ARPEGE, NEMO, OASIS3-MCT
EC-EARTH3 SMHI IFS, NEMO, OASIS3
HadGEM3 UK Metoffice GA6.0, GL6.0, NEMO, CICE, OASIS3
CESM-NEMO
CESM-NEMO CMCC CAM, NEMO, CLM, CICE, RTM, CPL7
NorESM MET.no CAM-Oslo, NCC-MICOM, CLM, CICE, CPL7

The software infrastructure for M4HR experiments does not only include the actual ESMs but also further components needed to complete the workflow. The list of software components that receive special attention are:

Component nameInstitutionCompnent type
OASIS CERFACS Coupler
XIOS CNRS-IPSL I/O subsystem
CDI-PIO DKRZ I/O subsystem
CDO MPG Postrpocessing/Data analysis tool
Autosubmit IC3 Job control tool
Cylc UREAD-NCAS Job control tool
Rose MetOfiice Job control tool

Metrics

In contrast to tradional performance metrics (such as the rate of floating point operations (FLOPS)) the metrics chosen for the M4HR experiment are chosen to reflect the typical questions ESM users have when they plan an experiment run. Such as:

  • How long will the experiment take (including data transfer and post-processing)?
  • How many nodes (cores+memory) can be efficiently used in different phases of the experiment?
  • Are there bottlenecks in the experiment workflow?
  • How much short-term/medium-term/long-term disk space is needed?
  • Can/should the experiment be split up in parallel chunks (e.g. How many ensemble members should be run in parallel?)?

 

The performance metrics are based on the ideas of V. Balaji, as presented at the IS-ENES Workshop “Exascale Technologies and Innovation in HPC for Climate Models” and were adjusted and extended to fit the needs of the M4HR experiment. The complete list of used metrics can be found here.

Results

Initial performance analysis results and a more detailed description of the experimental setup can be found in the IS-ENES2 deliverable D9.1.