Is your feature request related to a problem? Please describe.
In my transient cases I would like to average the flow-field over many time-iterations, in order to filter out highly dynamic effects. For me the volumentric time-averages and rms values have significant value. However, I have observed, that once averaging is enabled, the solver performance gradually decreases up to a point, where time-iterations are orders of magnitudes slower than without averaging. This does not seem to be a bug but rather inherent to the way averaging is implemented.

I suspect, that the averaging occurs over all timesteps and thus a complete history of all the values is required. This may gradually increase memory consumption and effort for calculating averages. I suspect this to happen around here, but I may be wrong.

I assume, that this may be a necessity for the higher-order windows, but seems to be overkill for a rectangular window, where the new time-average can be computed from the previous time-average, the number of iterations and the new values. Nevertheless, also the rectangular window seems to save the full value history:

Describe the solution you'd like
Being able to calculate (at least) non-windowed time-averages without a gradual drop in performance would be a real asset for me. This may be possible by choosing a different averaging method. If this cannot be done due to the data structure used, a dry-run message warning the user about this may be helpful.

Describe alternatives you've considered
The same effect can be achieved by running simulations without time-averaging but with regular volumetric outputs and calculating time-averages as a post-processing step. However, the associated data overhead is considerable.

Additional context
I performed my tests so far with the RANS solver, employing SST turbulence and SLAU2 scheme. Time-averaging started at around time-iteration 160k and cintinous for another 200k iterations.