virtual memory 4472

I remember similar conclusions (a) from when I was at Gould 1985-1989, where we put the latest and greatest working set scheduler into UNIX - and then backed off (b) from reading the literature but I'll rely on others to provide references.

Methinks this is the virtual memory paging analogy of what I posted about with respect to hardware caches some time ago: everybody talks about parbreastioned LRU, but it almost never pays off. Global LRU, and various approximations are where we are at.

In that earlier post about caches I ventured the guess that, if some sort of local replacement is to be a win, it must be mixed with global replacement. Since determining cache demand is hard, I suggested randomization - e.g. use local 9-10s of the time, global 1-10 of the time. Randomized to prevent resonance.

I conjecture that the same thing may apply to virtual memory paging systems. Although paging systems have much more time available to applies "smarts" to determine memory demand, I don't know how well these work for a varying workload. Randomized mixing of local and global working set policies may help.

Or maybe local just sucks irretrievably.