Author tries to optimize for execution time (finish program fast). But most users want to decrease total CPU resource consumption, because most tasks runs in parallel on server. In second case traditional hotspot profiling works just fine.
That can definitely be true, but note lots of things are latency-, not throughput-bound. In particular, microservice-based systems are likely to hit 99% latency of some service on (many/most/all) requests; I'd expect that a similar trick would be helpful there.
