Challenge 31 - CLAM - Cloud-computing based template for NWP Local Area Model #8
Description
Stream 3 - Software Development
Mentors
- Roberto Cuccu
- Bojan Kasic
- Alon Shtivelman
Skill Required
- Experience with Linux system and scripting in Linux environment.
- Knowledge of compiling and running scientific codes.
- Experience with installation and running of one NWP Local Area Model would be desirable. Since this project aims to explore configurations of NWP model (domain size and resolution) which can be run on resources available in the European Weather Cloud (EWC), any open-source model (ICON, WRF, etc.) used in national weather services of ECMWF Member States (MS) or Co-operating States (CS) is welcome to be used in this project.
- Familiarity with automation tools such as Ansible applied in a cloud-computing platform context would be desired.
- ECMWF can provide support for work in EWC environment, support with compiling and limited support with model configuration and running depending on selected model. However, it is preferable that potential candidates have previous experience with the model used in this project.
Goal
Testing and benchmarking of NWP Local Area Model on the European Weather Cloud (EWC) and delivery of a template that other EWC users can re-use and customise for their own activity.
Investigating limits in terms of domain size and resolution which can be run in reasonable time on resources available on EWC.
Description of the Challenge
As part of its mandate ECMWF manages computing facilities providing vital resources for producing global numerical weather predictions and other data for our Member and Co-operating States and the broader community. These facilities include a high-performance computing facility (HPCF), which relatively recently has been complemented with a cloud computing infrastructure which powers the ECMWF side of the European Weather Cloud (EWC), a joint initiative from ECMWF and EUMETSAT providing a cloud-computing-based collaboration platform for the use of the meteorological community.
According to the ECMWF convention, access to the computing facilities is made available to the Member States (MS) and Co-operating States (CS) for their research and duties. This usage from MS/CS users in some cases includes also the execution of NWP local area model to produce weather forecasting on a custom regional domain. Despite the execution of these LAM models best fits the use of HPC facility, there are some use cases that could benefit from a flexible resource provisioning on the cloud infrastructure to run these models on demand. This might be especially relevant for users that by default do not have full access to the ECMWF's HPC (e.g. Co-operating States users).
In this context, the European Weather Cloud (EWC) could provide a suitable environment for deploying and running the limited area NWP models and address the needs from the use cases that could take advantage of such cloud-computing platform. It promotes the sharing of resources and collaboration across the community and makes available also a dedicated space, the EWC Community-Hub, where users share items that could be useful to others on the platform.
This Code4Earth challenge aims at addressing the following goals:
- feasibility assessment to run an NWP Local Area Model on EWC Virtual Machines.
- benchmarking the execution of these models according to different configuration (e.g. by varying the extent of the selected domain and resolution) in relation to the employed EWC Virtual Machine resources.
- setup of a reusable template that could be submitted as community contribution to the EWC Community-Hub.
What is the current problem / limitation?
Users of the European Weather Cloud (EWC) may want to run a NWP LAM model on the available resources (Virtual Machines) and would be facilitated if there is a ready-to-use template that allows them to setup the environment in few steps. This Code4Earth challenge aims at assessing the feasibility and providing this reusable template by making it available to interested users.
For example, ECMWF's Co-operating States do not have access to ECMWF's HPC and have limited human resources for setting up NWP model in a cloud environment on their own. They could greatly benefit from results of this project.
What data / system do you plan to use?
[Data]
A limited area NWP model can be initialised with data produced by ECMWF IFS global model. ECMWF OpenData: https://www.ecmwf.int/en/forecasts/datasets/open-data already contains mandatory fields required to initialise some models like WRF.
If the selected model cannot be initialised with OpenData content, ECMWF can provide a sample dataset from IFS model which contains additional parameters or another GRIB version which are not available in OpenData.
Work will be done on EWC.
[NWP LAM Model]
Challenge participants may choose to propose and work with a NWP limited area model that is available as open source, which may include for instance:
WRF (https://github.com/wrf-model/WRF)
ICON (https://gitlab.dkrz.de/icon/icon-model)
or any other open source NWP model used by Co-operating States users and the community.
The selection of the model to adopt on the challenge should take into consideration the existing user uptake of such model in order to maximise the number of potential interested users that could benefit of the outcome of this challenge. Ideally, the model should be used in some MS or CS national weather services.
[EWC cloud resources]
As part of the challenge, it is beneficial to test and benchmark the model runs on different flavours of Virtual Machines (VM) listed in the EWC - ECMWF Plans documentation. Access to the cloud resources will be made available to the participant by the mentors of the challenge.
At least initially, as a basic setup the model deployment and runs should be conducted on a single Virtual Machine. More complex setups could be also proposed and considered in addition to the basic setup.
The number of VM resource plans where to tests and benchmark the model execution should be determined with the objective to provide the end-user with a sufficient estimation of the overall model performance and limitations according to the different VM resources configuration.
The available VM images and Operating Systems versions are listed in the dedicated VM images page on the EWC documentation.
What could be the solution?
The desired outcome of the challenge is to build a template and made available for the EWC community so that any user of the EWC can :
- deploy the Local Area Model template on the EWC with minimum effort and selecting the required cloud resources.
- configure custom parameters for the Local Area Model (e.g. area of interest, resolution, etc).
- perform several test runs with different configurations.
if time permits, we could run the model on-demand or in scheduled mode but this is not mandatory.
Ideas for implementation
Possible steps to be followed can be:
- Define an initial development/testing environment (e.g. target Virtual Machine plan and Operating System).
- Check the required dependencies to run the selected model (e.g. compilers, software packages and tools, ..) in relation to the target environment.
- Make sure the dependencies are satisfied by automating the installation and configuration where required and possible.
- Deploy the selected NWP model on the target environment.
- Run tests and benchmarks by varying model configuration and employed cloud resources.
- Collect and track the benchmarking results.
- Automate the model deployment (e.g. using Ansible) creating a template which should minimize the user interaction to have an environment as much as ready to use.
- Document the steps for the template usage.
- Prepare a contribution to the EWC Community-Hub [https://confluence.ecmwf.int/x/JQVbIg] according to the guidelines.
Additional considerations and ideas:
- The model once tested would typically be executed in unattended mode for a scheduled production. As an added value, the template could includes a mechanism to setup a time-based scheduled run of the model.
- The use of containerized model deployment might be considered.
- For the template packaging it is possible to take as examples the available items contribution on the EWC Community-Hub for what concerning the deployment automation and the technologies adopted.
Evaluation criteria
- Feasibility
- Easy to maintain / Future-proof approach
- Transferability
- Comprehensibility