Re: Fixed at last!

Am 01.06.2012 13:25, schrieb Nigel Cunningham:
> Hi all.
>
> I thought you might like to hear some good news.
>
> I finally found some time to put a concerted effort into finding and
> fixing the cause of TuxOnIce not working with the BFS scheduler. The
> issue turned out, I think, to be the same one that's been causing those
> pesky occasional failures to read compressed data (the -22 error).
>
> The problem turned out to be nice and simple: when we create the extra
> threads for doing multithreaded I/O, we bind each one to its own CPU.
> That's find and dandy until we get to the atomic copy, when we do CPU
> hotplugging. The affinities get broken and threads end up stomping over
> each others' memory. It seems that BFS does a better job of breaking the
> affinities than the normal scheduler, which made the problem easier to
> find.
>
> So, after thinking all this time that it was a locking problem in the
> code that was already there, it turned out to be a few lines of code
> that weren't there instead: the solution is to simply get each thread to
> remember its affinity and reset it before it starts to do any work.

Great work, thanks a lot! Finally fixed and now we can use tuxonice with 
BFS again.

Harald

-- 
`Experience is the best teacher.'