On Thu, 2013-09-12 at 19:17 -0400, Mikulas Patocka wrote: > > On Thu, 12 Sep 2013, Mike Snitzer wrote: > > > On Thu, Sep 12 2013 at 6:47pm -0400, > > Mikulas Patocka <mpatocka@xxxxxxxxxx> wrote: > > > > > We don't need this. bio-based i/o should be fine with a small mempool, > > > there is no need to make it tunable. > > > > I'd like to get Frank's insight here. He clearly had an interest in > > tuning bio-based also. While 16 shouldn't really hurt it could still be > > artifically high. I'm not opposed to exposing a sane default but > > allowing other to experiemnt (in production workloads) with smaller > > values that still enable forward progress should memory get exhausted. > > > > Mike > > I would do it this way: if Frank gets a measurable improvement in memory > consumption when the values is dropped from 16 to a lower number (4 or > maybe 1), then I would drop the value by default (don't make it tunable, > drop it for all users). > > If there is no improvement when the value is lowered, I'd leave it as it > is, on 16. I haven't had time to look at this lately, but I did see a small but measurable reduction in memory consumption when I was dealing with this before. The real problem for us is the fact that we'll have tons of these devices so even a small memory reduction means a lot when multiplied by, say, a hundred devices across a thousand machines. And that's a _small_ number. That said, I wouldn't want to see all Linux installs limited by our specific case, which is why I would vote for a tunable rather than a fixed, hardcoded value. -- Frank Mayhar 310-460-4042 -- dm-devel mailing list dm-devel@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/dm-devel
