André Braga <meianoite@xxxxxxxxx> wrote: > Then the load-balancer migrates this thread to the best destination > core among the candidates. Let W be the quantity of cores considered > for receiving migrated threads, so that K <= W <= 2K. This is so that > the algorithm scales reasonably well for a large number of cores. Not sure this hog array will work so well - even lots of non-hogs hurt latency which might be worse depending on what you want to achieve. A load is balanced when each CPU has a similar thread profile (ie. the ratio between interactive and computing threads), even though this might not be the best solution if throughput is your worry. > In order to avoid the ping-pong effect, a just-migrated thread could > have a "cool-down" period where it would not be allowed to get to the > "hogs" list. If there would be a ping-pong effect, I guess the load balancer wouldn't do a good job :-) Bye, Axel.