On 2017年12月21日 16:17, Michal Hocko wrote:
> On Thu 21-12-17 16:06:50, kemi wrote:
>>
>>
>> On 2017年12月20日 18:12, Michal Hocko wrote:
>>> On Wed 20-12-17 13:52:14, kemi wrote:
>>>>
>>>>
>>>> On 2017年12月19日 20:40, Michal Hocko wrote:
>>>>> On Tue 19-12-17 14:39:24, Kemi Wang wrote:
>>>>>> We have seen significant overhead in cache bouncing caused by NUMA
>>>>>> counters
>>>>>> update in multi-threaded page allocation. See 'commit 1d90ca897cb0 ("mm:
>>>>>> update NUMA counter threshold size")' for more details.
>>>>>>
>>>>>> This patch updates NUMA counters to a fixed size of (MAX_S16 - 2) and
>>>>>> deals
>>>>>> with global counter update using different threshold size for node page
>>>>>> stats.
>>>>>
>>>>> Again, no numbers.
>>>>
>>>> Compare to vanilla kernel, I don't think it has performance improvement, so
>>>> I didn't post performance data here.
>>>> But, if you would like to see performance gain from enlarging threshold
>>>> size
>>>> for NUMA stats (compare to the first patch), I will do that later.
>>>
>>> Please do. I would also like to hear _why_ all counters cannot simply
>>> behave same. In other words why we cannot simply increase
>>> stat_threshold? Maybe calculate_normal_threshold needs a better scaling
>>> for larger machines.
>>>
>>
>> I will add this performance data to changelog in V3 patch series.
>>
>> Test machine: 2-sockets skylake platform (112 CPUs, 62G RAM)
>> Benchmark: page_bench03
>> Description: 112 threads do single page allocation/deallocation in parallel.
>> before after
>> (enlarge threshold size)
>> CPU cycles 722 379(-47.5%)
>
> Please describe the numbers some more. Is this an average?
Yes
> What is the std?
I increase the loop times to 10m, so the std is quite slow (repeat 3 times)
> Can you see any difference with a more generic workload?
>
I didn't see obvious improvement for will-it-scale.page_fault1
Two reasons for that:
1) too long code path
2) server zone lock and lru lock contention (access to buddy system frequently)
>> Some thinking about that:
>> a) the overhead due to cache bouncing caused by NUMA counter update in fast
>> path
>> severely increase with more and more CPUs cores
>
> What is an effect on a smaller system with fewer CPUs?
>
Several CPU cycles can be saved using single thread for that.
>> b) AFAIK, the typical usage scenario (similar at least)for which this
>> optimization can
>> benefit is 10/40G NIC used in high-speed data center network of cloud
>> service providers.
>
> I would expect those would disable the numa accounting altogether.
>
Yes, but it is still worthy to do some optimization, isn't?
