Re: [RFC PATCH 0/14] Parallel memory initialisation

[nzimmer <nzimmer@sgi.com>]

On 04/15/2015 10:44 AM, Mel Gorman wrote:
> On Wed, Apr 15, 2015 at 10:50:45AM -0400, Waiman Long wrote:
>> On 04/15/2015 09:38 AM, Mel Gorman wrote:
>>> On Wed, Apr 15, 2015 at 09:15:50AM -0400, Waiman Long wrote:
>>>>> <SNIP>
>>>>> Patches are against 4.0-rc7.
>>>>>
>>>>>   Documentation/kernel-parameters.txt |   8 +
>>>>>   arch/ia64/mm/numa.c                 |  19 +-
>>>>>   arch/x86/Kconfig                    |   2 +
>>>>>   include/linux/memblock.h            |  18 ++
>>>>>   include/linux/mm.h                  |   8 +-
>>>>>   include/linux/mmzone.h              |  37 +++-
>>>>>   init/main.c                         |   1 +
>>>>>   mm/Kconfig                          |  29 +++
>>>>>   mm/bootmem.c                        |   6 +-
>>>>>   mm/internal.h                       |  23 ++-
>>>>>   mm/memblock.c                       |  34 ++-
>>>>>   mm/mm_init.c                        |   9 +-
>>>>>   mm/nobootmem.c                      |   7 +-
>>>>>   mm/page_alloc.c                     | 398 +++++++++++++++++++++++++++++++-----
>>>>>   mm/vmscan.c                         |   6 +-
>>>>>   15 files changed, 507 insertions(+), 98 deletions(-)
>>>>>
>>>> I had included your patch with the 4.0 kernel and booted up a
>>>> 16-socket 12-TB machine. I measured the elapsed time from the elilo
>>>> prompt to the availability of ssh login. Without the patch, the
>>>> bootup time was 404s. It was reduced to 298s with the patch. So
>>>> there was about 100s reduction in bootup time (1/4 of the total).
>>>>
>>> Cool, thanks for testing. Would you be able to state if this is really
>>> important or not? Does booting 100s second faster on a 12TB machine really
>>> matter? I can then add that justification to the changelog to avoid a
>>> conversation with Andrew that goes something like
>>>
>>> Andrew: Why are we doing this?
>>> Mel:    Because we can and apparently people might want it.
>>> Andrew: What's the maintenance cost of this?
>>> Mel:    Magic beans
>>>
>>> I prefer talking to Andrew when it's harder to predict what he'll say.
>> Booting 100s faster is certainly something that is nice to have.
>> Right now, more time is spent in the firmware POST portion of the
>> bootup process than in the OS boot.
> I'm not surprised. On two different 1TB machines, I've seen a post time
> of 2 minutes and one of 35. No idea what it's doing for 35 minutes....
> plotting world domination probably.
>
>> So I would say this patch isn't
>> really critical right now as machines with that much memory are
>> relatively rare. However, if we look forward to the near future,
>> some new memory technology like persistent memory is coming and
>> machines with large amount of memory (whether persistent or not)
>> will become more common. This patch will certainly be useful if we
>> look forward into the future.
>>
> Whether persistent memory needs struct pages or not is up in the air and
> I'm not getting stuck in that can of worms. 100 seconds off kernel init
> time is a starting point. I can try pushing it on on that basis but I
> really would like to see SGI and Intel people also chime in on how it
> affects their really large machines.
>
I will get some numbers from this patch set but I haven't had the 
opportunity yet.  I will grab them this weekend for sure if I can't get 
machine time sooner.


>>>> However, there were 2 bootup problems in the dmesg log that needed
>>>> to be addressed.
>>>> 1. There were 2 vmalloc allocation failures:
>>>> [    2.284686] vmalloc: allocation failure, allocated 16578404352 of
>>>> 17179873280 bytes
>>>> [   10.399938] vmalloc: allocation failure, allocated 7970922496 of
>>>> 8589938688 bytes
>>>>
>>>> 2. There were 2 soft lockup warnings:
>>>> [   57.319453] NMI watchdog: BUG: soft lockup - CPU#1 stuck for 23s!
>>>> [swapper/0:1]
>>>> [   85.409263] NMI watchdog: BUG: soft lockup - CPU#1 stuck for 22s!
>>>> [swapper/0:1]
>>>>
>>>> Once those problems are fixed, the patch should be in a pretty good
>>>> shape. I have attached the dmesg log for your reference.
>>>>
>>> The obvious conclusion is that initialising 1G per node is not enough for
>>> really large machines. Can you try this on top? It's untested but should
>>> work. The low value was chosen because it happened to work and I wanted
>>> to get test coverage on common hardware but broke is broke.
>>>
>>> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
>>> index f2c96d02662f..6b3bec304e35 100644
>>> --- a/mm/page_alloc.c
>>> +++ b/mm/page_alloc.c
>>> @@ -276,9 +276,9 @@ static inline bool update_defer_init(pg_data_t *pgdat,
>>>   	if (pgdat->first_deferred_pfn != ULONG_MAX)
>>>   		return false;
>>>
>>> -	/* Initialise at least 1G per zone */
>>> +	/* Initialise at least 32G per node */
>>>   	(*nr_initialised)++;
>>> -	if (*nr_initialised>  (1UL<<  (30 - PAGE_SHIFT))&&
>>> +	if (*nr_initialised>  (32UL<<  (30 - PAGE_SHIFT))&&
>>>   	(pfn&  (PAGES_PER_SECTION - 1)) == 0) {
>>>   		pgdat->first_deferred_pfn = pfn;
>>>   		return false;
>> I will try this out when I can get hold of the 12-TB machine again.
>>
> Thanks.
>
>> The vmalloc allocation failures were for the following hash tables:
>> - Dentry cache hash table entries
>> - Inode-cache hash table entries
>>
>> Those hash tables scale linearly with the amount of memory available
>> in the system. So instead of hardcoding a certain value, why don't
>> we make it a certain % of the total memory but bottomed out to 1G at
>> the low end?
>>
> Because then it becomes what percentage is the right percentage and what
> happens if it's a percentage of total memory but the NUMA nodes are not
> all the same size?. I want to start simple until there is more data on
> what these really large machines look like and if it ever fails in the
> field, there is the command-line switch until a patch is available.
>