System Requirements for Installing and Running WRF

[kwerner]

We understand the desire for specific figures. However, providing exact numbers is challenging without considering factors such as domain sizes, vertical levels, physical parameterizations, and output frequency. Therefore, we can only offer general recommendations at this time.

Regarding system requirements for installing and running WRF, all components of the system should be considered.

• WRF system requirements: WPS executables use minimal memory unless the domain is very large (thousands x thousands of grid cells) and can run serially or with MPI. Ungrib, which decodes grib files independently of the WRF domain, must be built and run serially; even large global datasets fit on laptop memory.
• The real program uses more memory than WRFl.
• WRF simulation time is typically semi-equivalent to post-processing time.

Consider the following factors when selecting an appropriate machine:

• If the machine will primarily run production tasks with distributed memory jobs, the amount of memory per node can be reduced. If the machine will be used for a combination of distributed memory jobs (which aggregate memory across multiple nodes) and single processor jobs (such as post-processors and visualization), then memory should likely be increased.
• The necessary machine size and core count depend on the job type. Time-sensitive forecasts require larger, potentially underutilized machines. A mix of small and large independent jobs benefits from smaller, continuously running machines. Running multiple independent jobs, like ensembles, is more efficient on several multi-processor machines than a single machine with the same total cores.
• The WRF model does not take advantage of GPU, Xeon Phi, or any other accelerator technology. If the machine is purchased primarily for WRF, there is no need to include accelerators in your purchase. For a multi-purpose machine with use of graphics and visualization, having GPU-populated login nodes may be beneficial.

Recommendations:

1. WRF support primarily assists with Unix/Linux systems using GNU and Intel compilers, as these align with NCAR's experience on Intel hardware. Virtual environments often present challenges. While direct access to supercomputer architectures like Fujitsu and Cray is limited, vendors offer user support. Adhering to these common configurations allows for more effective assistance.
2. In distributed memory systems, prioritize increasing the number of processors over the amount of memory, as memory can be aggregated.
3. For heterogeneous machines, ensure the login nodes (master node) are well-equipped with memory.
4. The WRF model can output data from individual processors. Due to the significant communication volume it requires, high bandwidth between processors and to I/O systems is essential for optimal performance.
5. A few desktop machines networked with Ethernet cables are not an effective cluster. Dedicated networking infrastructure is recommended.
6. Disk space is relatively inexpensive. If the machine will be utilized for analysis, a few TB of disk will not be sufficient.

After several months of using a new machine, users typically develop clear preferences and opinions about their purchase, identifying both positive and negative aspects. These experiences and insights from a larger user group can be valuable to consider when evaluating new equipment.