On Thu, Dec 12, 2013 at 01:58:59PM -0500, Brian Foster wrote:
> On 12/12/2013 02:22 AM, Dave Chinner wrote:
> > From: Dave Chinner <dchinner@xxxxxxxxxx>
> >
> > The per-ag locks used to protect per-ag block lists are located in a tightly
> > packed array. That means that they share cachelines, so separate them out
> > inot
> > separate 64 byte regions in the array.
> >
> > pahole confirms the padding is correctly applied:
> >
> > struct aglock {
> > pthread_mutex_t lock; /* 0 40 */
> >
> > /* size: 64, cachelines: 1, members: 1 */
> > /* padding: 24 */
> > };
> >
> > Signed-off-by: Dave Chinner <dchinner@xxxxxxxxxx>
> > ---
>
> Out of curiosity, any data on this one?
There's a small improvement but it's within the margins of error. At
higher levels of concurrency it will make a bigger difference, but
right now we can't get to a thread per AG (where the contention
would really show) because of other scalability limitations
e.g. mmap_sem contention becomes a limiting factor at 20-25 threads
because of page faults occurring on hash cache lookups (read lock)
occurring in parallel with memory allocation (libc mprotect calls
take a write lock). So more scalability will come from a CPU cache
friendlier cache index - I'm tempted to drop in per-AG rb trees and
see where that leads (now where would I have got that idea?)....
Cheers,
Dave.
--
Dave Chinner
david@xxxxxxxxxxxxx
|