gitpeterwind
commented
7 months ago Hi Ruben, This may well be the case. The reason is (probably) that it is not that simple to know in advance which portion of the map is needed on each MPI domain (the 1km map is in polar stereographic projection). Since we do not need the 1km map in most of our runs, we have not prioritized to improve this. Maybe if you only use a small section of this map, you could try to make your own smaller map that cover only your region?
Best wishes, Peter
rubenww
mifads
Dear EMEP modellers,
We've noticed that the initial memory usage during the reading in of the land use maps is very high in EMEP and that it scales with the number of cpu cores that is used. I'm using v4.45 currently.
In our configuration, it turned out that we needed approximately 1.5GB / cpu core in memory, or EMEP would crash during the reading in of the land use map, after the following steps:
As you can see, we're using the 1x1km landuse file. During the rest of the run, only approximately 250 MB / cpu core in memory is needed for the calculation. This is more in line with the requirements stated in the EMEP user guide. The domain size is 153x135 cells in this case (approx. 3x6 km resolution, lon-lat).
I briefly looked in the code in ReadLanduse_CDF from Landuse_mod.f90 (in EMEP v4.45) and noticed the following comment:
If indeed the landuse is read in for each processor, this would explain the high initial memory usage. Is that indeed the case and if so, are there any fundamental reasons why this is necessary?
Or could the code perhaps be adjusted in such a way that the same domain decomposition is applied to the land use maps, so that the memory requirements for running EMEP on multiple cores can be lower?
To illustrate, we're typically using 24 cores and this requires approximately 34GB in memory due to the above "problem", while after the land use has been read in, only around 6 GB of memory is in use.
Thank you for your help.
Best, Ruben Verweij