Re: [PATCH] mem: balanced allocation of hugepages

[Ilya Maximets <i.maximets@samsung.com>]

On 07.04.2017 18:44, Thomas Monjalon wrote:
> 2017-04-07 18:14, Ilya Maximets:
>> Hi All.
>>
>> I wanted to ask (just to clarify current status):
>> Will this patch be included in current release (acked by maintainer)
>> and then I will upgrade it to hybrid logic or I will just prepare v3
>> with hybrid logic for 17.08 ?
> 
> What is your preferred option Ilya?

I have no strong opinion on this. One thought is that it could be
nice if someone else could test this functionality with current
release before enabling it by default in 17.08.

Tomorrow I'm going on vacation. So I'll post rebased version today
(there are few fuzzes with current master) and you with Sergio may
decide what to do.

Best regards, Ilya Maximets.

> Sergio?
> 
> 
>> On 27.03.2017 17:43, Ilya Maximets wrote:
>>> On 27.03.2017 16:01, Sergio Gonzalez Monroy wrote:
>>>> On 09/03/2017 12:57, Ilya Maximets wrote:
>>>>> On 08.03.2017 16:46, Sergio Gonzalez Monroy wrote:
>>>>>> Hi Ilya,
>>>>>>
>>>>>> I have done similar tests and as you already pointed out, 'numactl --interleave' does not seem to work as expected.
>>>>>> I have also checked that the issue can be reproduced with quota limit on hugetlbfs mount point.
>>>>>>
>>>>>> I would be inclined towards *adding libnuma as dependency* to DPDK to make memory allocation a bit more reliable.
>>>>>>
>>>>>> Currently at a high level regarding hugepages per numa node:
>>>>>> 1) Try to map all free hugepages. The total number of mapped hugepages depends if there were any limits, such as cgroups or quota in mount point.
>>>>>> 2) Find out numa node of each hugepage.
>>>>>> 3) Check if we have enough hugepages for requested memory in each numa socket/node.
>>>>>>
>>>>>> Using libnuma we could try to allocate hugepages per numa:
>>>>>> 1) Try to map as many hugepages from numa 0.
>>>>>> 2) Check if we have enough hugepages for requested memory in numa 0.
>>>>>> 3) Try to map as many hugepages from numa 1.
>>>>>> 4) Check if we have enough hugepages for requested memory in numa 1.
>>>>>>
>>>>>> This approach would improve failing scenarios caused by limits but It would still not fix issues regarding non-contiguous hugepages (worst case each hugepage is a memseg).
>>>>>> The non-contiguous hugepages issues are not as critical now that mempools can span over multiple memsegs/hugepages, but it is still a problem for any other library requiring big chunks of memory.
>>>>>>
>>>>>> Potentially if we were to add an option such as 'iommu-only' when all devices are bound to vfio-pci, we could have a reliable way to allocate hugepages by just requesting the number of pages from each numa.
>>>>>>
>>>>>> Thoughts?
>>>>> Hi Sergio,
>>>>>
>>>>> Thanks for your attention to this.
>>>>>
>>>>> For now, as we have some issues with non-contiguous
>>>>> hugepages, I'm thinking about following hybrid schema:
>>>>> 1) Allocate essential hugepages:
>>>>>     1.1) Allocate as many hugepages from numa N to
>>>>>          only fit requested memory for this numa.
>>>>>     1.2) repeat 1.1 for all numa nodes.
>>>>> 2) Try to map all remaining free hugepages in a round-robin
>>>>>     fashion like in this patch.
>>>>> 3) Sort pages and choose the most suitable.
>>>>>
>>>>> This solution should decrease number of issues connected with
>>>>> non-contiguous memory.
>>>>
>>>> Sorry for late reply, I was hoping for more comments from the community.
>>>>
>>>> IMHO this should be default behavior, which means no config option and libnuma as EAL dependency.
>>>> I think your proposal is good, could you consider implementing such approach on next release?
>>>
>>> Sure, I can implement this for 17.08 release.
>>>
>>>>>
>>>>>> On 06/03/2017 09:34, Ilya Maximets wrote:
>>>>>>> Hi all.
>>>>>>>
>>>>>>> So, what about this change?
>>>>>>>
>>>>>>> Best regards, Ilya Maximets.
>>>>>>>
>>>>>>> On 16.02.2017 16:01, Ilya Maximets wrote:
>>>>>>>> Currently EAL allocates hugepages one by one not paying
>>>>>>>> attention from which NUMA node allocation was done.
>>>>>>>>
>>>>>>>> Such behaviour leads to allocation failure if number of
>>>>>>>> available hugepages for application limited by cgroups
>>>>>>>> or hugetlbfs and memory requested not only from the first
>>>>>>>> socket.
>>>>>>>>
>>>>>>>> Example:
>>>>>>>>      # 90 x 1GB hugepages availavle in a system
>>>>>>>>
>>>>>>>>      cgcreate -g hugetlb:/test
>>>>>>>>      # Limit to 32GB of hugepages
>>>>>>>>      cgset -r hugetlb.1GB.limit_in_bytes=34359738368 test
>>>>>>>>      # Request 4GB from each of 2 sockets
>>>>>>>>      cgexec -g hugetlb:test testpmd --socket-mem=4096,4096 ...
>>>>>>>>
>>>>>>>>      EAL: SIGBUS: Cannot mmap more hugepages of size 1024 MB
>>>>>>>>      EAL: 32 not 90 hugepages of size 1024 MB allocated
>>>>>>>>      EAL: Not enough memory available on socket 1!
>>>>>>>>           Requested: 4096MB, available: 0MB
>>>>>>>>      PANIC in rte_eal_init():
>>>>>>>>      Cannot init memory
>>>>>>>>
>>>>>>>>      This happens beacause all allocated pages are
>>>>>>>>      on socket 0.
>>>>>>>>
>>>>>>>> Fix this issue by setting mempolicy MPOL_PREFERRED for each
>>>>>>>> hugepage to one of requested nodes in a round-robin fashion.
>>>>>>>> In this case all allocated pages will be fairly distributed
>>>>>>>> between all requested nodes.
>>>>>>>>
>>>>>>>> New config option RTE_LIBRTE_EAL_NUMA_AWARE_HUGEPAGES
>>>>>>>> introduced and disabled by default because of external
>>>>>>>> dependency from libnuma.
>>>>>>>>
>>>>>>>> Cc:<stable@dpdk.org>
>>>>>>>> Fixes: 77988fc08dc5 ("mem: fix allocating all free hugepages")
>>>>>>>>
>>>>>>>> Signed-off-by: Ilya Maximets<i.maximets@samsung.com>
>>>>>>>> ---
>>>>>>>>    config/common_base                       |  1 +
>>>>>>>>    lib/librte_eal/Makefile                  |  4 ++
>>>>>>>>    lib/librte_eal/linuxapp/eal/eal_memory.c | 66 ++++++++++++++++++++++++++++++++
>>>>>>>>    mk/rte.app.mk                            |  3 ++
>>>>>>>>    4 files changed, 74 insertions(+)
>>>>
>>>> Acked-by: Sergio Gonzalez Monroy <sergio.gonzalez.monroy@intel.com>
>>>
>>> Thanks.
>>>
>>> Best regards, Ilya Maximets.
>>>
> 
> 
> 
> 
>