Interview: Andrea Arcangeli

Posted May 19, 2004 22:02 UTC (Wed) by jzbiciak (✭ supporter ✭, #5246) [Link]

Not a big upheaval, really. As Andrea notes in the interview, the VM heuristics are not changing this time, but rather how you identify a page once you've picked it. It's a much simpler piece that's changing. The impact of the change is mostly in memory footprint required for the VM algorithms. 'rmap' and 'objrmap' are both reverse-mapping techniques--they just achieve the same goal (tracing a physical page to all its virtual mappings) by different bookkeeping. Which pages get swapped shouldn't change dramatically between the two approaches, although objrmap should result in significantly less memory pressure on big iron.

In contrast, the 2.4 upheaval changed the fundamental heuristics by which the VM picks pages to push out to swap. That's a much bigger deal. As I understand it, the old 2.4 algorithm was a clock-based algorithm that scaled poorly to a large number of pages. The AA algorithm was more of a global LRU. The cost of identifying a page to swap out, in addition to the actual swapping behavior under load can be quite different with the two approaches. It's truly a much more fundamental change.