On Tue, Aug 09, 2005 at 09:15:26PM +0200, Peter Zijlstra wrote:

On Tue, 2005-08-09 at 15:25 -0300, Marcelo Tosatti wrote:

quoted

Hi Rik,

Two hopefully useful comments:

i) ARC and its variants requires additional information about page
replacement (namely whether the page has been reclaimed from the L1 or
L2 lists).

How costly would it be to add this information to the hash table?

I've been thinking on reserving another word in the cache-line and use
that as a bit-array to keep that information; the only problems with
that would be atomicy of the {bucket,bit} tuple and very large
cachelines where NUM_NR > 32. 

The chance for a lookup hit to happen on a hash value which is in a
modified-state in a different CPU's cacheline should be pretty small
(depends on the architecture also, but shouldnt be much of an issue I
guess).

Hoping on that, guaranteed validity of data is not necessary, it is OK
to be incorrect occasionally.

quoted

ii) From my reading of the patch, the provided "distance" information is
relative to each hash bucket. I'm unable to understand the distance metric
being useful if measured per-hash-bucket instead of globally?

The assumption is that IFF the hash function has good distribution
properties the per bucket distance is a good approximation of
(distance >> nonres_shift).

Well, not really "good approximation" it sounds to me, the sensibility
goes down to L1_CACHE_LINE/sizeof(u32), which is:

- 8 on 32-byte cacheline
- 16 on 64-byte cacheline 
- 32 on 128-byte cacheline

Right?

So the (nice!) refault histogram gets limited to those values?

quoted

PS: Since remember_page() is always called with the zone->lru_lock held,
the preempt_disable/enable pair is unecessary at the moment... still, 
might be better to leave it there for safety reasons.

There being multiple zones; owning zone->lru_lock does not guarantee
uniqueness on the remember_page() path as its a global structure.

True, but it guarantees disabled preemption. No big deal...
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>