> > > I'm surprised this ack a bit. Rik, do you have any improvement plan about > > > streaming io detection logic? > > > I think the patch have a slightly marginal benefit, it help to<1% scan > > > ratio case. but it have big regression, it cause streaming io (e.g. backup > > > operation) makes tons swap. > > > > How? From the description I believe it took 16GB in > > a zone before we start scanning anon pages when > > reclaiming at DEF_PRIORITY? > > > > Would that casue a problem? > > Please remember, 2.6.27 has following +1 scanning modifier. > > zone->nr_scan_active += (zone_page_state(zone, NR_ACTIVE) >> priority) + 1; > ^^^^ > > and, early (ano not yet merged) SplitLRU VM has similar +1. likes > > scan = zone_nr_lru_pages(zone, sc, l); > scan >>= priority; > scan = (scan * percent[file]) / 100 + 1; > ^^^ > > We didn't think only one page scanning is not big matter. but it was not > correct. we got streaming io bug report. the above +1 makes annoying swap > io. because some server need big backup operation rather much much than > physical memory size. > > example, If vm are dropping 1TB use once pages, 0.1% anon scanning makes > 1GB scan. and almost server only have some gigabyte swap although it > has >1TB memory. > > If my memory is not correct, please correct me. > > My point is, greater or smaller than 16GB isn't essential. all patches > should have big worth than the downside. The description said "the impact > sounds not a big deal", nobody disagree it. but it's worth is more little. > I don't imagine this patch improve anything. And now I've merged this patch into my local vmscan patch queue. After solving streaming io issue, I'll put it to mainline. Thanks. -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxxx For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>