"Aneesh Kumar K.V" <aneesh.kumar@xxxxxxxxxxxxx> writes:
> "Huang, Ying" <ying.huang@xxxxxxxxx> writes:
>
>> Aneesh Kumar K V <aneesh.kumar@xxxxxxxxxxxxx> writes:
>>
>>> On 7/26/22 1:14 PM, Huang, Ying wrote:
>>>> "Aneesh Kumar K.V" <aneesh.kumar@xxxxxxxxxxxxx> writes:
>>>>
>
> ....
>
>>>> + */
>>>>> +int next_demotion_node(int node)
>>>>> +{
>>>>> + struct demotion_nodes *nd;
>>>>> + int target;
>>>>> +
>>>>> + if (!node_demotion)
>>>>> + return NUMA_NO_NODE;
>>>>> +
>>>>> + nd = &node_demotion[node];
>>>>> +
>>>>> + /*
>>>>> + * node_demotion[] is updated without excluding this
>>>>> + * function from running.
>>>>> + *
>>>>> + * Make sure to use RCU over entire code blocks if
>>>>> + * node_demotion[] reads need to be consistent.
>>>>> + */
>>>>> + rcu_read_lock();
>>>>> + /*
>>>>> + * If there are multiple target nodes, just select one
>>>>> + * target node randomly.
>>>>> + *
>>>>> + * In addition, we can also use round-robin to select
>>>>> + * target node, but we should introduce another variable
>>>>> + * for node_demotion[] to record last selected target node,
>>>>> + * that may cause cache ping-pong due to the changing of
>>>>> + * last target node. Or introducing per-cpu data to avoid
>>>>> + * caching issue, which seems more complicated. So selecting
>>>>> + * target node randomly seems better until now.
>>>>> + */
>>>>> + target = node_random(&nd->preferred);
>>>>
>>>> In one of the most common cases, nodes_weight(&nd->preferred) == 1.
>>>> Where, get_random_int() in node_random() just wastes CPU cycles and
>>>> random entropy. So the original struct demotion_nodes implementation
>>>> appears better.
>>>>
>>>> struct demotion_nodes {
>>>> unsigned short nr;
>>>> short nodes[DEMOTION_TARGET_NODES];
>>>> };
>>>>
>>>
>>>
>>> Is that measurable difference? using nodemask_t makes it much easier with respect to
>>> implementation. IMHO if we observe the usage of node_random() to have performance impact
>>> with nodes_weight() == 1 we should fix node_random() to handle that? If you strongly
>>> feel we should fix this, i can opencode node_random to special case node_weight() == 1?
>>
>> If there's no much difference, why not just use the existing code?
>> IMHO, it's your responsibility to prove your new implementation is
>> better via numbers, for example, reduced code lines, with better or same
>> performance.
>>
>> Another policy is just to use the existing code in the first version.
>> Then change it based on measurement.
>
> One of the reason I switched to nodemask_t is to make code simpler.
> demotion target is essentially a node mask.
>
>>
>> In general, I care more about the most common cases, that is, 0 or 1
>> demotion target.
>
> How about I switch to the below opencoded version. That should take care
> of the above concern.
Per my estimation, the performance for 0 or 1 demotion target should be
OK.
And I think that you can change node_random() implementation directly.
Because it will not hurt other users too.
Best Regards,
Huang, Ying
>>
>>> - target = node_random(&nd->preferred);
>>> + node_weight = nodes_weight(nd->preferred);
>>> + switch (node_weight) {
>>> + case 0:
>>> + target = NUMA_NO_NODE;
>>> + break;
>>> + case 1:
>>> + target = first_node(nd->preferred);
>>> + break;
>>> + default:
>>> + target = bitmap_ord_to_pos(nd->preferred.bits,
>>> + get_random_int() % node_weight, MAX_NUMNODES);
>>> + break;
>>> + }
>>>
>>>
