thats afaik not done because of the performance overhead.

most requests take only a fraction of a second to answer. to actually monitor threads on that frequency you'd need not only a significant portion for the monitoring (ever looked at the CPU usage with running top?) but also for the network communication to send that over to the lb, which needs to process the data for load balancing

You wouldn't need to monitor individual threads and processes. Only average response time for each server or total CPU load.