RFR: 8272807: Permit use of memory concurrent with pretouch
kbarrett at openjdk.java.net
Tue Aug 24 07:33:28 UTC 2021
On Mon, 23 Aug 2021 15:59:47 GMT, Aleksey Shipilev <shade at openjdk.org> wrote:
> > Performance testing hasn't found any regressions related to this change.
> Are there any performance tests that actually check `-XX:+AlwaysPreTouch` performance? It is very likely that the OS trip to commit the page is vastly costlier than the atomic increment per page, but it would be nice to confirm.
I've run various configurations of some of our "usual" performance benchmarks
with and without this change and haven't seen any difference at all. (Changing
the value of AlwaysPreTouch does have an effect on some, with the right other
configuration options.) Even with the work I've been doing to move some of the
pretouching to make use of the new feature, it can be hard to find any difference.
I *haven't* written a microbenchmark to commit memory, time touching it with
one of the approaches, uncommit, repeat. I could do that, though I don't
expect it to show anything either.
More information about the hotspot-runtime-dev