LHC@home

Your strategy seems to be to reduce the task run time by assigning more CPUs to tasks. That can work but only if the host has enough memory.
The following app_config.xml should work to run max 1 ATLAS vbox task at a time with 2 CPUs per task if the host has enough memory.
If it works then you could try 2 tasks concurrently by increasing the <max_concurrent> (not <project_max_concurrent>) value to 2.

The <project_max_concurrent> is commented out between <!-- and --> tags because it is optional.
Some of the other lines are optional too but I am not sure how those work. I think (maybe) <avg_ncpus> overrides the <cmdline>--nthreads</cmdline> which would mean you can eliminate the <cmdline>--nthreads</cmdline> if you include the <avg_ncpus>. Or maybe it's the opposite. If both lines are included then their values should match. As far as I know it does no harm to include both.

<app_config>
   <!-- <project_max_concurrent>4</project_max_concurrent> -->
   <app>
     <name>ATLAS</name>
     <max_concurrent>1</max_concurrent>
   </app>
   <app_version>
     <app_name>ATLAS</app_name>
     <plan_class>vbox64_mt_mcore_atlas</plan_class>
     <avg_ncpus>2</avg_ncpus>
     <cmdline>--nthreads 2 </cmdline>
     <!-- RAM formula: RAM = 3000 + 900 * ncpus -->
     <cmdline>--memory_size_mb 4800 </cmdline>
   </app_version>
</app_config>

Are you still looking for a method to get smaller tasks?
As bronco already explained, there is no method.

The only thing you can do is to change the basic settings (#cpus, RAM size) already given by the server.
This makes sense if the basic settings lead to computation errors or make your computer sluggish.
Otherwise it's recommended to leave them untouched.

If you want to receive fewer tasks (for less remaining time), then just reduce the size of the buffer in BOINC Computing preferences.
I use the default of 0.1 + 0.5 days.

Okey, thank you! I'll try to change this parameters.

That will get you fewer tasks but they will not be smaller tasks.
All ATLAS task consist of 200 events. There is no way to get ATLAS tasks that have less than 200 events. The time required to process the 200 varies and is not predictable.

If you want to receive fewer tasks (for less remaining time), then just reduce the size of the buffer in BOINC Computing preferences.
I use the default of 0.1 + 0.5 days.

HM, I had already made a note about this:

[Update 18.09.2018] Nowerdays Atlas runs only MultiCoreWUs, even if you run it with 1-Core only it will need up to 3,9 GB as SingleCore

To change the web preferences is the easiest way to get a good setup as it also adjusts the RAM calculation.
The server sends the correct values within the scheduler_reply.
This method is recommended if you run only 1 subproject, e.g. ATLAS.

#cpus and RAM settings from the server can be overruled by a local app_config.xml.
If you use "<avg_ncpus>x</avg_ncpus>", "<cmdline>--nthreads ..." should be obsolete.

BTW:
It's "nthreads" instead of "nthreats"

But i don't understand why my single-core tasks consume more than 2,1 GB of memory RAM, as indicated in Yeti's Checklist ...

It was necessary to change the RAM fromula a couple of times during the last years.
Yeti also posted the recent one:
Yeti wrote:

... MultiCore-WUs need 3,0 GB + 0,9 GB * number of cores

I had some similar happenings. The only but working solution was to run Atlas-WUs as Single-Core and this worked fine

Some rare comments you can find with google point out that the error may be a result of bad physical RAM.
Could you check this with memtest?