Local Bags? about crossbeam HOT 5 CLOSED

I think trying to get local bags to work since using only the global GC in it's current (and in my opinion, any) incarnation seems pretty bad for performance

I agree with this, for exactly the reasons you mention -- it's why I added local bags originally.

My decision to revert to global bags was to try to get back to a conservative, known-good state for the library as quickly as possible. That, and the fact that the local bag code was a little smelly any, so I'd rather re-add a fresh version.

re: having four epochs, I agree that this seems the simplest approach to recovering local bags, and should address the bug I'm worried about.

I'll tidy up my test code and add it to the repo ASAP. Thanks!

from crossbeam.

I think trying to get local bags to work since using only the global GC in it's current (and in my opinion, any) incarnation seems pretty bad for performance

I definitely wrote that sentence too late at night!

My decision to revert to global bags was to try to get back to a conservative, known-good state for the library as quickly as possible. That, and the fact that the local bag code was a little smelly any, so I'd rather re-add a fresh version.

I totally agree, especially at this stage of the project. Tonight I'll test out the 4-bag version and see how it goes. I'll add any interesting test cases I find as well.

from crossbeam.

FWIW, the average push/pop time for MsQueue on my machine is ~120ns with local bags and ~160ns with only global bags.

from crossbeam.

Now that @msullivan's fix has landed, we should work on reincorporating local bags. As I said previously, the local bag code was due for a refactor anyway, so I'd be happy to start fresh. And I do think we can get away with the 3 epoch version, based on what you and @msullivan argued.

from crossbeam.

Thanks @msullivan for #51!!

from crossbeam.