Today's mainframeanything to new 554

That could have to do with some aspect of the tty support unrelated to load. It would be interesting to profile the code and see exactly where it was during this delay.

My experience with Windows on 8 way systems running large DBMS or web services loads has been that "feel" under load is somewhat unpredictable. Sometimes it's in proportion to load and other times it gets very bursty with long periods of unresponsiveness. Task scheduling under Windows seems to be an issue.

I once ran an Oracle DBMS server on a 4 way HP K580 with 4GB RAM under HP-UX 10.20 with over 1,000 users doing largely sales invoicing and ad hoc reporting (not mbuttive reporting jobs, which we batched for offpeak runs). The "feel" of this box was quite snappy even when load average reached 27 or so, and DBMS request time never exceeded about 950ms. The reason was use of HP-UX realtime scheduler rather than the default nice scheduler. The box was near its limit, but degradation was very smooth with no deadlocks or bursty delays.

Of course it depends on the type of tasks running, but I think much depends on task scheduling in a real world workload.

"Feel" is also dependent on the terminal interface code, which can have odd effects depending on buffering, character handling, etc. which may be independent of system load but may be exposed under load.