:orphan: ============== Compiling FAQs ============== | | WRF/WPS Computation Requirements ================================ * WPS processes use little memory. Though geogrid and metgrid support distributed-memory/parallel processing, even for a 1000x1000 grid cell domain, they can typically run on a single processor. * Ungrib, a program for degribbing GRIB meteorological data, operates independently of the WRF domain. It handles large 2D arrays (e.g., 1440x720 for 0.25-degree global data) that fit easily on laptop memory. Ungrib must be run serially, regardless of the WPS build. * The real program, the vertical-level pre-processor to the WRF model, uses more memory than the WRF model. * WRF model simulation and post-processing often take similar amounts of time. | Consider the following for machine appropriateness: * For production tasks with distributed memory jobs, memory per node can be reduced. For a mix of distributed-memory (which can aggregate memory across multiple nodes) and single-processor jobs (e.g., post-processing and visualization), memory likely needs to be increased. * Machine configuration (cores per node, total nodes per machine, etc.) depends on machine requirements. For semi-operational, time-sensitive forecasts, a larger, though not always fully utilized, machine is needed. For a mixture of small and large production jobs (e.g., a 1-year simulation), a smaller, 24/7 machine might suffice. For independent, multiple jobs (e.g., ensembles), several smaller n-processor machines are more efficient than one large machine with the same total cores. * The WRF model does not utilize GPU, xeon phi, or any other accelerator technology. For machines primarily for WRF, accelerators are not necessary. For multi-purpose machines, including graphics and visualization, GPU-populated login nodes may be beneficial. | Recommendations --------------- .. note:: We are unable to provide specifics. Due to numerous factors that go into simulations (e.g., domain size, vertical levels, physics options, output frequency), we can only provide general recommendations. | #. We cannot test all compiler, chip, and OS combinations. We have extensive experience with Intel chips and Unix/Linux operating systems, and regularly test GNU and Intel compilers. Adhering to these norms allows us to provide better assistance with WRF. While we don't have direct access to supercomputer architectures like Fujitsu or Cray, their vendors offer us direct user support. Users with virtual environments often encounter more issues. #. For distributed-memory machines, money is better spent increasing the number of processors than the amount of memory (since memory can be aggregated). #. For heterogeneous machines, fill the login/master node with memory. #. Because WRF can output data from each processor, it requires a fairly large amount of communication. Bandwidth between processors and bandwidth to IO systems is critical. #. Purchasing dedicated networking is essential. Simply connecting a few desktop machines with ethernet cables is not an ideal cluster. #. Disk space is relatively inexpensive. For machines utilized for analysis, only a few TB of disk space is insufficient. | It may be beneficial to gauge reviews from others who are content with their setup. Typically only a few months is necessary to form opinions on the pros and cons of users’ system purchases. | | | Configuration Options ===================== | When configuring WRF, a list of processing types is listed. See `Configure WRF `_ for descriptions of the processing choices. Alternative to a standard configuration (``./configure``), `see Configure WRF `_ for other available options for debugging and precision. `See Configure WPS `_ or WPS-specific configuration options. | | | | Installing WRF and WPS ====================== Follow `Complete Installation of WRF and WPS `_ for prerequisite, computing environment setup, library installation, and the steps to build WRF and WPS. | | | | |