Back again, with some more info but unfortunately no solution :-|

First of all, thank you for your continued support. I guess you're doing this in your free time. If you know of any "official" support channel for WCG, please let me know what those are, since I feel bad for using your time.

Now back to the issue at hand:

- it seems to continue. It has now been 10 days since the cluster size increase, and while "Results returned" have stabilized at the new level, "Points Generated" and "Total Run Time" have still not changed (so this seems to indicate that it isn't an issue with a delay in the finished WUs being validated).

- my device installations are constantly increasing (because the instances in the cluster are cleaned up by Google Cloud after 24 hours and replaced with new ones).

- I prefer WUs for MCM and SCC, though I have checked "If there is no work available for the project(s) I have selected above, please send me work from another project." in my profile. So the machines are are working mainly on MCM WUs

- I logged in to a couple of such machines and looked at "boinc.log". From that it seems that it's finishing WUs (after ~20 hours of processing) and uploading them.

- I looked at the "My contribution" page and things look "normal" - as far as I can tell. The instances all points generated / results returned. I also tried to check what the "points generated" / "results returned" is and for all instances it's between 500 and 800 (this is the same range my laptop is in for example).

- On the results page I have no "Invalid" results. I do have a lot of results in the "Error" state where it says Status: Detached, but I guess that's normal since it happens whenever an preemtible instance is cleaned up.

My next step is to update the "global_prefs_override.xml" for the cluster so that it downloads less WUs in advance (it doesn't make sense because the instances have very good network connection all the time and hopefully this also will reduce the "detached" WUs). Particularly:

<suspend_cpu_usage>80.000000</suspend_cpu_usage>
<work_buf_min_days>0.000000</work_buf_min_days>
<work_buf_additional_days>0.000000</work_buf_additional_days>
<max_ncpus_pct>100.000000</max_ncpus_pct>
<disk_max_used_gb>0.000000</disk_max_used_gb>
<disk_max_used_pct>100.000000</disk_max_used_pct>
<disk_min_free_gb>0.000000</disk_min_free_gb>
<cpu_usage_limit>100.000000</cpu_usage_limit>


Also, I'm adding the following to cc_config.xml to perhaps improve the computation speed a little:

<lower_client_priority>0</lower_client_priority>
<no_priority_change>1</no_priority_change>

I also looked into solutions to persist state between restarts of preemtible VMs, but I didn't find anything straight-forward (my biggest stumbling block is the following: lets say I have a shared storage with the structure /instances/boinc-1, /instances/boinc-2, ...etc - how does an instance know when it comes up which directory to use?)

Again, these changes probably will improve the situation somewhat (a little bit more processing time and less detached WUs), but somehow I don't think it will solve the underlying issue.

I keep hammering away at this because I don't want all that processing power to be spent in wain...