Oh, another idea I had at the conference that I just remembered. A useful thing -- when the system's CPU usage drops below 10%, it sends a SIGBORED to each process in turn, and if they catch it instead of ignoring it, it waits for the CPU usage to drop again below 10% again, and continues to send them. Once all processes get it, it then waits awhile before doing it again. Default handler is to ignore. There might be a process, kidled, that asks the kernel to do things that are best handled in kernel-space. Why? So the system can do housekeeping when it's not otherwise busy. Why not defrag the disk? Have Perl or your JVM do garbage-collection? Have Postgres flush it's dead table entries. Etc etc.