Re: [RFC] high system time & lock contention running large mixed workload

[KOSAKI Motohiro]

> On Mon, Nov 30, 2009 at 05:00:29PM -0500, Larry Woodman wrote:
> > Before the splitLRU patch shrink_active_list() would only call
> > page_referenced() when reclaim_mapped got set.  reclaim_mapped only got
> > set when the priority worked its way from 12 all the way to 7. This
> > prevented page_referenced() from being called from shrink_active_list()
> > until the system was really struggling to reclaim memory.
> 
> page_referenced should never be called and nobody should touch ptes
> until priority went down to 7. This is a regression in splitLRU that
> should be fixed. With light VM pressure we should never touch ptes ever.

Ummm. I can't agree this. 7 is too small priority. if large system have prio==7,
the system have unacceptable big latency trouble.
if only prio==DEF_PRIOTIRY or something, I can agree you probably.


> > On way to prevent this is to change page_check_address() to execute a
> > spin_trylock(ptl) when it was called by shrink_active_list() and simply
> > fail if it could not get the pte_lockptr spinlock.  This will make
> > shrink_active_list() consider the page not referenced and allow the
> > anon_vma->lock to be dropped much quicker.
> > 
> > The attached patch does just that, thoughts???
> 
> Just stop calling page_referenced there...
> 
> Even if we ignore the above, one problem later in skipping over the PT
> lock, is also to assume the page is not referenced when it actually
> is, so it won't be activated again when page_referenced is called
> again to move the page back in the active list... Not the end of the
> world to lose a young bit sometime though.
> 
> There may be all reasons in the world why we have to mess with ptes
> when there's light VM pressure, for whatever terabyte machine or
> whatever workload that performs better that way, but I know in 100% of
> my systems I don't ever want the VM to touch ptes when there's light
> VM pressure, no matter what. So if you want the default to be messing
> with ptes, just give me a sysctl knob to let me run faster.

Um.
Avoiding lock contention on light VM pressure is important than
strict lru order. I guess we don't need knob.


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/