high memory usage when appending `uds` to `partitions` list

[veenstrajelmer]

Running the following script called memory_usage.py with memory_profiler via mprof run python memory_usage.py and mprof plot:

import os
import glob
import xugrid as xu
import xarray as xr
import datetime as dt
from time import sleep

def open_part_ds(file_nc_list):
    print(f'>> xu.open_dataset() with {len(file_nc_list)} partition(s): ',end='')
    dtstart = dt.datetime.now()
    partitions = []
    for iF, file_nc_one in enumerate(file_nc_list):
        print(iF+1,end=' ')
        ds_one = xr.open_mfdataset(file_nc_one, chunks="auto")
        uds_one = xu.core.wrap.UgridDataset(ds_one)
        partitions.append(uds_one)
    print(': ',end='')
    print(f'{(dt.datetime.now()-dtstart).total_seconds():.2f} sec')

dir_model = r"p:\11210284-011-nose-c-cycling\runs_fine_grid\B05_waq_2012_PCO2_ChlC_NPCratios_DenWat_stats_2023.01\B05_waq_2012_PCO2_ChlC_NPCratios_DenWat_stats_2023.01\DFM_OUTPUT_DCSM-FM_0_5nm_waq"
file_nc_pat = os.path.join(dir_model, "DCSM-FM_0_5nm_waq_0*_map.nc")
file_nc_list_all = glob.glob(file_nc_pat)
file_nc_list = file_nc_list_all[:5]

uds = open_part_ds(file_nc_list)
sleep(2)

Results in this memory usage:

However, when commenting partitions.append(uds_one), we get way less memory usage and we see garbage collection in action:

The accumulating memory consumption upon appending is inconvenient, since we want to make a list of partitions for xu.merge_partitions(). When calling gc.collect() after xr.open_dataset() (or elsewhere), this does not make a difference.

Might be related to:

• https://github.com/pydata/xarray/issues/2186

[veenstrajelmer]

What is very interesting is that using a with statement resolves all this. This time including merging:

import os
import glob
import xugrid as xu
import xarray as xr
import datetime as dt
from time import sleep

def open_part_ds(file_nc_list, withwith):
    print(f'>> xu.open_dataset() with {len(file_nc_list)} partition(s): ',end='')
    dtstart = dt.datetime.now()
    partitions = []
    for iF, file_nc_one in enumerate(file_nc_list):
        print(iF+1,end=' ')
        if withwith:
            with xr.open_mfdataset(file_nc_one, chunks="auto") as ds_one:
                uds_one = xu.core.wrap.UgridDataset(ds_one)
                partitions.append(uds_one)
        else:
            ds_one = xr.open_mfdataset(file_nc_one, chunks="auto")
            uds_one = xu.core.wrap.UgridDataset(ds_one)
            # ds_one.close()
            # uds_one.close()
        partitions.append(uds_one)
    print(': ',end='')
    print(f'{(dt.datetime.now()-dtstart).total_seconds():.2f} sec')

    print('>> xu.merge_partitions(): ',end='')
    dtstart = dt.datetime.now()
    uds = xu.merge_partitions(partitions)
    print(f'{(dt.datetime.now()-dtstart).total_seconds():.2f} sec')
    return uds

dir_model = r"p:\11210284-011-nose-c-cycling\runs_fine_grid\B05_waq_2012_PCO2_ChlC_NPCratios_DenWat_stats_2023.01\B05_waq_2012_PCO2_ChlC_NPCratios_DenWat_stats_2023.01\DFM_OUTPUT_DCSM-FM_0_5nm_waq"
file_nc_pat = os.path.join(dir_model, "DCSM-FM_0_5nm_waq_0*_map.nc")
file_nc_list_all = glob.glob(file_nc_pat)
file_nc_list = file_nc_list_all[:5]

uds = open_part_ds(file_nc_list, withwith=False)
sleep(2)

withwith=False:

withwith=True:

Or withwith=False and ds.close() (uds.close() does not do the trick):

From this it can be concluded that it is wise to close the original xarray dataset if not using it anymore. The time/memory consumption by merging will be unaffected by this. I will at least pick this up in https://github.com/Deltares/dfm_tools/issues/968, but it might also be good to add it to the xugrid documentation. Adding it to xu.open_dataset() has no added benefit, since users might use the returned uds partition directly (e.g. for removing ghost cells or so), and in that case the memory consumption will be back.

[veenstrajelmer]

If the user also does another action (like plotting a single timestep) on the merged dataset, the memory usage increases again to the usage that we saw without ds.close(). This is documented in https://github.com/Deltares/dfm_tools/issues/968 and a clean version in https://github.com/Deltares/dfm_tools/issues/484. Therefore, it seems not useful to close the datasets after all. Furthermore, it is clear that engine="h5netcdf" consumes way less memory (40 MB instead of 110 MB per partition), but xr.open_dataset() showed to be way slower for datasets with many variables like in this example. This might be fixed by https://github.com/h5netcdf/h5netcdf/issues/195.

[veenstrajelmer]

Since this is not an issue with xugrid, this issue can be closed.