(2011/07/13 20:02), Mel Gorman wrote:
> On Wed, Jul 13, 2011 at 10:15:15AM +0900, KOSAKI Motohiro wrote:
>> (2011/07/11 22:01), Mel Gorman wrote:
>>> The zonelist cache (ZLC) is used among other things to record if
>>> zone_reclaim() failed for a particular zone recently. The intention
>>> is to avoid a high cost scanning extremely long zonelists or scanning
>>> within the zone uselessly.
>>>
>>> Currently the zonelist cache is setup only after the first zone has
>>> been considered and zone_reclaim() has been called. The objective was
>>> to avoid a costly setup but zone_reclaim is itself quite expensive. If
>>> it is failing regularly such as the first eligible zone having mostly
>>> mapped pages, the cost in scanning and allocation stalls is far higher
>>> than the ZLC initialisation step.
>>>
>>> This patch initialises ZLC before the first eligible zone calls
>>> zone_reclaim(). Once initialised, it is checked whether the zone
>>> failed zone_reclaim recently. If it has, the zone is skipped. As the
>>> first zone is now being checked, additional care has to be taken about
>>> zones marked full. A zone can be marked "full" because it should not
>>> have enough unmapped pages for zone_reclaim but this is excessive as
>>> direct reclaim or kswapd may succeed where zone_reclaim fails. Only
>>> mark zones "full" after zone_reclaim fails if it failed to reclaim
>>> enough pages after scanning.
>>>
>>> Signed-off-by: Mel Gorman <mgorman@xxxxxxx>
>>
>> If I understand correctly this patch's procs/cons is,
>>
>> pros.
>> 1) faster when zone reclaim doesn't work effectively
>>
>
> Yes.
>
>> cons.
>> 2) slower when zone reclaim is off
>
> How is it slower with zone_reclaim off?
>
> Before
>
> if (zone_reclaim_mode == 0)
> goto this_zone_full;
> ...
> this_zone_full:
> if (NUMA_BUILD)
> zlc_mark_zone_full(zonelist, z);
> if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
> ...
> }
>
> After
> if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
> ...
> }
> if (zone_reclaim_mode == 0)
> goto this_zone_full;
> this_zone_full:
> if (NUMA_BUILD)
> zlc_mark_zone_full(zonelist, z);
>
> Bear in mind that if the watermarks are met on the first zone, the zlc
> setup does not occur.
Right you are. thank you correct me.
>> 3) slower when zone recliam works effectively
>>
>
> Marginally slower. It's now calling zlc setup so once a second it's
> zeroing a bitmap and calling zlc_zone_worth_trying() on the first
> zone testing a bit on a cache-hot structure.
>
> As the ineffective case can be triggered by a simple cp, I think the
> cost is justified. Can you think of a better way of doing this?
So, now I'm revisit your number in [0/3]. and I've conclude your patch
improve simple cp case too. then please forget my last mail. this patch
looks nicer.
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@xxxxxxxxxxxxxx>
>
>> (2) and (3) are frequently happen than (1), correct?
>
> Yes. I'd still expect zone_reclaim to be off on the majority of
> machines and even when enabled, I think it's relatively rare we hit the
> case where the workload is regularly falling over to the other node
> except in the case where it's a file server. Still, a cp is not to
> uncommon that the kernel should slow to a crawl as a result.
>
>> At least, I think we need to keep zero impact when zone reclaim mode is off.
>>
>
> I agree with this but I'm missing where we are taking the big hit with
> zone_reclaim==0.
>
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>