On Thursday 20 December 2007 18:04, Christoph Lameter wrote:
> > The only reason the x86 ticket locks have the 256 CPu limit is that
> > if they go any bigger, we can't use the partial registers so would
> > have to have a few more instructions.
>
> x86_64 is going up to 4k or 16k cpus soon for our new hardware.
>
> > A 32 bit spinlock would allow 64K cpus (ticket lock has 2 counters,
> > each would be 16 bits). And it would actually shrink the spinlock in
> > the case of preempt kernels too (because it would no longer have the
> > lockbreak field).
> >
> > And yes, I'll go out on a limb and say that 64k CPUs ought to be
> > enough for anyone ;)
>
> I think those things need a timeframe applied to it. Thats likely
> going to be true for the next 3 years (optimistic assessment ;-)).
Yeah, that was tongue in cheek ;)
> Could you go to 32bit spinlock by default?
On x86, the size of the ticket locks is 32 bit, simply because I didn't
want to risk possible alignment bugs (a subsequent patch cuts it down to
16 bits, but this is a much smaller win than 64->32 in general because
of natural alignment of types).
Note that the ticket locks still support twice the number as the old
spinlocks, so I'm not causing a regression here... but yes, increasing
the size further will require an extra instruction or two.
> How about NUMA awareness for the spinlocks? Larger backoff periods for
> off node lock contentions please.
ticket locks can naturally tell you how many waiters there are, and how
many waiters are in front of you, so it is really nice for doing backoff
(eg. you can adapt the backoff *very* nicely depending on how many are in
front of you, and how quickly you are moving toward the front).
Also, since I got rid of the ->break_lock field, you could use that space
perhaps to add a cpu # of the lock holder for even more backoff context
(if you find that helps).
Anyway, I didn't do any of that because it obviously needs someone with
real hardware in order to tune it properly.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[Index of Archives]
[Kernel Newbies]
[Netfilter]
[Bugtraq]
[Photo]
[Stuff]
[Gimp]
[Yosemite News]
[MIPS Linux]
[ARM Linux]
[Linux Security]
[Linux RAID]
[Video 4 Linux]
[Linux for the blind]
[Linux Resources]
