Problems with collect () or take (n) on MB sized data by spark

We are a group of people working with spark (spark 2.1.0 standalone with 2 workers; programming is done in scala and everything runs inside multiple dockers). We are faced with a problem that "collect" or "receive (n)" is very slow when the collected data hits some cap size.

We ran into the problem several times, but we boiled down to a simple example: it reads a file (either from the local filesystem or hdfs; we tested both) and then the result. It works fine up to a certain file size (about 2 MB) and then it is very slow (and about 3 MB it breaks completely). If it doesn't compile (for example, it just does saveAtextFile), the install can deal with files up to 200 GB in size. We tested the increase (from 2 GB of RAM to 20 GB of RAM), but this did not solve the problem; in fact, our tests show that our little experiment slows down at the same file size no matter how much RAM we give the driver or workers.

I've summarized my experiment here: https://pastebin.com/raw/6yXztq0H

In this experiment, the program reads file "s" and "take (n)" with "n" incrementally increases. As the timestamped result shows, it works almost instantaneously for "n≤104145" (it changes almost a little, despite the big changes in the setting), and then it is quite slow. For large "n" (see second run), the driver fails with "TaskResultLost" error. The last experiment (third run) shows that this does not seem to be a memory problem (and it seems logical since the file is relatively small, about 2 MB).

(This is not shown in the experiment, but we also played with SPARK_DAEMON_MEM, but that doesn't change anything.)

Has anyone had the same problem? Does anyone have an idea to help us look further?