Originally posted by archkde
View Post
The case where I think there is work still to be done here is in thread/pool/task/async type things. How you go from 0 workers to 1 worker (extremely common) can often be wasteful. (Do you wake up all threads saying I have a job for you? Do you round robin a fixed pool? Do you work steal? Do you create n_cpu threads at startup? Do you create a heartbeat thread? etc)
Anyway, this is why we should be profiling and using that profiler data to have meaningful conversations about software implementation.
Comment