Oh! I was not aware of the little cluster being that limited. Should be easy to check using taskset:
taskset -c 0-3 sysbench --test=cpu run --num-threads=4 --cpu-max-prime=20000
taskset -c 4-7 sysbench --test=cpu run --num-threads=4 --cpu-max-prime=20000
taskset -c 3 sysbench --test=cpu run --num-threads=1 --cpu-max-prime=20000
taskset -c 7 sysbench --test=cpu run --num-threads=1 --cpu-max-prime=20000
The numbers generated by the last 2 runs should be exactly four times larger than the two first runs.
Back then when you tested results with just 4 threads already showed a drop in performance. Curious…