Re: [PATCH] mm: do not drain pagevecs for mlock
From: KOSAKI Motohiro
Date: Fri Dec 30 2011 - 04:31:52 EST
2011/12/30 Tao Ma <tm@xxxxxx>:
> On 12/30/2011 04:11 PM, KOSAKI Motohiro wrote:
>> 2011/12/30 Tao Ma <tm@xxxxxx>:
>>> In our test of mlock, we have found some severe performance regression
>>> in it. Some more investigations show that mlocked is blocked heavily
>>> by lur_add_drain_all which calls schedule_on_each_cpu and flush the work
>>> queue which is very slower if we have several cpus.
>>> So we have tried 2 ways to solve it:
>>> 1. Add a per cpu counter for all the pagevecs so that we don't schedule
>>> and flush the lru_drain work if the cpu doesn't have any pagevecs(I
>>> have finished the codes already).
>>> 2. Remove the lru_add_drain_all.
>>> The first one has some problems since in our product system, all the cpus
>>> are busy, so I guess there is very little chance for a cpu to have 0 pagevecs
>>> except that you run several consecutive mlocks.
>>> From the commit log which added this function(8891d6da), it seems that we
>>> don't have to call it. So the 2nd one seems to be both easy and workable and
>>> comes this patch.
>> Could you please show us your system environment and benchmark programs?
>> Usually lru_drain_** is very fast than mlock() body because it makes
>> plenty memset(page).
> The system environment is: 16 core Xeon E5620. 24G memory.
> I have attached the program. It is very simple and just uses mlock/munlock.
Because your test program is too artificial. 20sec/100000times =
200usec. And your
program repeat mlock and munlock the exact same address. so, yes, if
lru_add_drain_all() is removed, it become near no-op. but it's
none of any practical program does such strange mlock usage.
But, 200usec is much than I measured before. I'll dig it a bit more.
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/