Before you submit a large simulation using the Multilevel Fast Multipole Method (MLFMM), use Feko’s built-in memory estimator to check whether the job is likely to fit on your workstation or cluster. The estimate takes a fraction of the time of a full solve and gives you a starting point for choosing a RAM allocation.
Run the memory estimate
1. In CADFEKO, open Solve/Run → Run/Launch → Feko Terminal.
2. From the model folder, run the command below. Replace my_model with the model’s base name and 8 with the number of processes you plan to use.
runfeko my_model -np 8 --feko-options --estimate-resource-requirements-only
3. When the estimate finishes, open the model’s .out file. Near the end, look for Memory estimate for a regular Feko run.
For a cluster job, run the estimate on the intended hosts with the same process allocation as the full solve. This example uses 40 processes across the two hosts defined in hosts.list.
runfeko my_model -np 40 --machines-file hosts.list --feko-options --estimate-resource-requirements-only
Running all 40 processes on a single node gives a less representative estimate for the two-node job because Feko duplicates some solution data by process or by host.
Read the right value in the output file
The .out file reports memory used to calculate the estimate alongside the predicted memory for the full solve. In the guide’s sample output, these differ substantially.
|
Output field |
Example value |
How to use it |
|
Peak memory usage during the whole solution |
61.707 MB |
Memory used by the master process during the estimation run in this parallel example. |
|
Sum of the peak memory of all processes |
1.928 GB |
Sum of process peaks during the estimation run. |
|
Memory estimate for a regular Feko run |
59.784 GB |
Use this total across all parallel processes as the starting point for planning the full solve. |
|
On average per process, under the full-run estimate |
1.868 GB |
An average across processes. Individual processes can use more. |
Tip
for cluster jobs
Check the RAM available on each node. Dividing the total estimate evenly across
nodes can understate the requirement on the busiest node.
Allow for memory beyond the estimate
• Output requests. Results such as far fields require additional memory that the estimator excludes. Leave headroom for the outputs in your production model.
• Mesh and preconditioner. The number of mesh elements, a mix of coarse and very fine regions, and the preconditioner’s type and size all affect memory. Feko supports estimation with both SPAI and sparse LU preconditioners.
• Parallel setup. With more processes, duplicated data can make the estimate less accurate. Repeat the estimate when you change the process count or host allocation.
• Memory errors. If the estimation run already fails with a memory allocation error, the full solve will also fail with that allocation. Increase the available RAM or revise the setup, then rerun the estimate.