Suspicious move evaluation speed with multi-threaded solving

[rsynek]

Please have a look at the following logs. Both come from running the same dataset for 10 minutes on CPUs with ARM architecture (full cores, no hyper-threading). The machine had 6 CPU cores available.

moveThreadCount=NONE

11:28:50.948  INFO Construction Heuristic phase (1) ended: time spent (165901), best score (0hard/-230000medium/-4740soft), move evaluation speed (29646/sec), step total (2364).
11:36:05.047  INFO Local Search phase (2) ended: time spent (600000), best score (0hard/-100000medium/-21933soft), move evaluation speed (10021/sec), step total (161167).
11:36:05.049  INFO Solving ended: time spent (600000), best score (0hard/-100000medium/-21933soft), move evaluation speed (15445/sec), phase total (3), environment mode (REPRODUCIBLE), move thread count (NONE).

moveThreadCount=5

11:12:53.382  INFO Construction Heuristic phase (1) ended: time spent (44150), best score (0hard/-230000medium/-4740soft), move evaluation speed (334507/sec), step total (2364).
11:22:09.232  INFO Local Search phase (2) ended: time spent (600000), best score (0hard/0medium/-5776soft), move evaluation speed (96346/sec), step total (566644).
11:22:09.235  INFO Solving ended: time spent (600001), best score (0hard/0medium/-5776soft), move evaluation speed (73973/sec), phase total (3), environment mode (REPRODUCIBLE), move thread count (5).

Observation 1: The move evaluation speed with 6 CPUs (1 for solver thread, 5 for move threads) is 10x higher to the speed with a single CPU. That is not proportional

Observation 2: for the multi-threaded run, the resulting speed is 73973/sec, which is below both the speed of CH and LS phases. there was also a custom phase running:

11:12:09.280 INFO Custom phase (0) ended: time spent (48), best score (0hard/-485066660000medium/0soft), move evaluation speed (0/sec), step total (1). ,

but given it took 48 ms, it should not impact the speed that much.