I took a extra long road though the application. 0.8 is more likely the upper baseline for the setup.
but see it mostly for benchmarking. its probably 10x as fast in production The interesting part is not that, but why worker mpm seems to torture the system beyond whats healthy when it should work faster side by side.

the total cpu was lower yes, and the %user part as well.
(i assume that would be apache). so on the paper it looks very cpu light and efficient, if i just could get my head around that system load.

locking is set 1100 times per request, locking + waiting time for one serial run is about 0.1s. in 10 concurrent i got a total (locking + waiting time) of 1.2s out of 28s per request. due to the long execution.

The benchmark is made from a simple perl script. spawning n threads with LWP::UserAgent request to the external server. clocking the request from fire to done. The target server is connected to the same local network so should not be any delay there.

tx for the tip about trying to isolate it to apache. i think it can be hard to do, its quite tied up to mod_perl with apache handlers. but i'll look into it

will also try to loose up the locks and see if that is contributing parameter here.