Date: Sat, 23 May 2015 11:49:33 +0800
From: Daniel J Blueman <daniel@numascale.com>
Subject: Re: [PATCH] mm: meminit: Finish initialisation of struct pages
 before basic setup
To: Waiman Long <waiman.long@hp.com>, Mel Gorman <mgorman@suse.de>
Cc: Andrew Morton <akpm@linux-foundation.org>, nzimmer <nzimmer@sgi.com>,
        Dave Hansen <dave.hansen@intel.com>,
        Scott Norton <scott.norton@hp.com>, Linux-MM <linux-mm@kvack.org>,
        LKML <linux-kernel@vger.kernel.org>,
        Steffen Persvold <sp@numascale.com>
Message-Id: <1432352973.6359.2@cpanel21.proisp.no>
In-Reply-To: <555F6404.4010905@hp.com>
References: <555F6404.4010905@hp.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8; format=flowed
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 3819
Lines: 95


-- 
Daniel J Blueman
Principal Software Engineer, Numascale

On Sat, May 23, 2015 at 1:14 AM, Waiman Long <waiman.long@hp.com> wrote:
> On 05/22/2015 05:33 AM, Mel Gorman wrote:
>> On Fri, May 22, 2015 at 02:30:01PM +0800, Daniel J Blueman wrote:
>>> On Thu, May 14, 2015 at 6:03 PM, Daniel J Blueman
>>> <daniel@numascale.com>  wrote:
>>>> On Thu, May 14, 2015 at 12:31 AM, Mel Gorman<mgorman@suse.de>  
>>>> wrote:
>>>>> On Wed, May 13, 2015 at 10:53:33AM -0500, nzimmer wrote:
>>>>>> I am just noticed a hang on my largest box.
>>>>>> I can only reproduce with large core counts, if I turn down the
>>>>>> number of cpus it doesn't have an issue.
>>>>>> 
>>>>> Odd. The number of core counts should make little a difference
>>>>> as only
>>>>> one CPU per node should be in use. Does sysrq+t give any
>>>>> indication how
>>>>> or where it is hanging?
>>>> I was seeing the same behaviour of 1000ms increasing to 5500ms
>>>> [1]; this suggests either lock contention or O(n) behaviour.
>>>> 
>>>> Nathan, can you check with this ordering of patches from Andrew's
>>>> cache [2]? I was getting hanging until I a found them all.
>>>> 
>>>> I'll follow up with timing data.
>>> 7TB over 216 NUMA nodes, 1728 cores, from kernel 4.0.4 load to 
>>> login:
>>> 
>>> 1. 2086s with patches 01-19 [1]
>>> 
>>> 2. 2026s adding "Take into account that large system caches scale
>>> linearly with memory", which has:
>>> min(2UL<<  (30 - PAGE_SHIFT), (pgdat->node_spanned_pages>>  3));
>>> 
>>> 3. 2442s fixing to:
>>> max(2UL<<  (30 - PAGE_SHIFT), (pgdat->node_spanned_pages>>  3));
>>> 
>>> 4. 2064s adjusting minimum and shift to:
>>> max(512UL<<  (20 - PAGE_SHIFT), (pgdat->node_spanned_pages>>  8));
>>> 
>>> 5. 1934s adjusting minimum and shift to:
>>> max(128UL<<  (20 - PAGE_SHIFT), (pgdat->node_spanned_pages>>  8));
>>> 
>>> 6. 930s #5 with the non-temporal PMD init patch I had earlier
>>> proposed (I'll pursue separately)
>>> 
>>> The scaling patch isn't in -mm.
>> That patch was superceded by "mm: meminit: finish
>> initialisation of struct pages before basic setup" and
>> "mm-meminit-finish-initialisation-of-struct-pages-before-basic-setup-fix"
>> so that's ok.
>> 
>> FWIW, I think you should still go ahead with the non-temporal 
>> patches because
>> there is potential benefit there other than the initialisation.  If 
>> there
>> was an arch-optional implementation of a non-termporal clear then it 
>> would
>> also be worth considering if __GFP_ZERO should use non-temporal 
>> stores.
>> At a greater stretch it would be worth considering if kswapd freeing 
>> should
>> zero pages to avoid a zero on the allocation side in the general 
>> case as
>> it would be more generally useful and a stepping stone towards what 
>> the
>> series "Sanitizing freed pages" attempts.

Good tip Mel; I'll take a look when time allows and get some data, 
though I guess it'll only be a win where the clearing is on a different 
node than the allocation.

> I think the non-temporal patch benefits mainly AMD systems. I have 
> tried the patch on both DragonHawk and it actually made it boot up a 
> little bit slower. I think the Intel optimized "rep stosb" 
> instruction (used in memset) is performing well. I had done similar 
> test on zero page code and the performance gain was non-conclusive.

I suspect 'rep stosb' on modern Intel hardware can write whole 
cachelines atomically, avoiding the RMW, or that the read part of the 
RMW is optimally prefetched. Open-coding it just can't reach the same 
level of pipeline saturation that the microcode can.

Daniel

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