Re: [PATCH] mm: do not drain pagevecs for mlock

From: Tao Ma
Date: Fri Jan 06 2012 - 01:47:33 EST


On 01/06/2012 02:33 PM, KOSAKI Motohiro wrote:
> (1/6/12 1:30 AM), Tao Ma wrote:
>> On 01/06/2012 02:18 PM, KOSAKI Motohiro wrote:
>>> 2012/1/6 Tao Ma<tm@xxxxxx>:
>>>> Hi Kosaki,
>>>> On 12/30/2011 06:07 PM, KOSAKI Motohiro wrote:
>>>>>>> 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
>>>>>>> worthless comparision.
>>>>>>> none of any practical program does such strange mlock usage.
>>>>>> yes, I should say it is artificial. But mlock did cause the
>>>>>> problem in
>>>>>> our product system and perf shows that the mlock uses the system time
>>>>>> much more than others. That's the reason we created this program
>>>>>> to test
>>>>>> whether mlock really sucks. And we compared the result with
>>>>>> rhel5(2.6.18) which runs much much faster.
>>>>>>
>>>>>> And from the commit log you described, we can remove
>>>>>> lru_add_drain_all
>>>>>> safely here, so why add it? At least removing it makes mlock much
>>>>>> faster
>>>>>> compared to the vanilla kernel.
>>>>>
>>>>> If we remove it, we lose to a test way of mlock. "Memlocked" field of
>>>>> /proc/meminfo
>>>>> show inaccurate number very easily. So, if 200usec is no avoidable,
>>>>> I'll ack you.
>>>>> But I'm not convinced yet.
>>>> Do you find something new for this?
>>>
>>> No.
>>>
>>> Or more exactly, 200usec is my calculation mistake. your program call
>>> mlock
>>> 3 times per each iteration. so, correct cost is 66usec.
>> yes, so mlock can do 15000/s, it is even slower than the whole i/o time
>> for some not very fast ssd disk and I don't think it is endurable. I
>> guess we should remove it, right? Or you have another other suggestion
>> that I can try for it?
>
> read whole thread.
I have read the whole thread, and you just described that the test case
is artificial and there is no suggestion or patch about how to resolve
it. As I have said that it is very time-consuming and with more cpu
cores, the more penalty, and an i/o time for a ssd can be faster than
it. So do you think 66 usec is OK for a memory operation?

Thanks
Tao
--
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/