I did some digging on the lkml forums (hard to add links from the bby) It looks like all of them relate to a series of commits for the kernel IPI process. It looks to be added as a way to load balance processes accross multiple cores in the fair scheduler (CFS) when the kernel is running tickless (config_no_hz). IPI is supposed to evaluate what core should handle the load balancing process across all cores at any given time, based on which ones are busy and idle. It looks to me to prefer keeping the load balance process on a lightly loaded core so that the heavier processes don't get impacted by the additional work of managing the other cores' processes. It sounds to me that it is a good way to keep high performance in a server (especially as you increase the number of cores in the system,) but I wonder if that may be the wong approach for a smaller system looking to maximize power efficiency instead of raw performance. On a single socket dual core system like atom, brazos, core2duo etc, it doesn't seem to be necessary as the cores on a single socket share cache and other resources. I wonder if it would make more sense in the smaller systems (single socket) to assign the load balancing process on one core, skip the IPI altogether and only wake the other cores when there is work to do. Perhaps even prefer to run new processes itself until it needs to wake another core. (Unfortunately, I am not a kernel hacker - If I am misunderstanding any of this, I would welcome someone to correct me :) On 6/10/11, Julian Kalinowski