rss-ladder and autorps/autoxps should warn about binding netdev to external NUMA node's CPUs

[strizhechenko]

rss-ladder eth1 1

while eth1 is NUMA0 bound should produce some warning about non-optimal performance.

• if there is no information about NUMA on given netdev there should be no warning
• if there is only one NUMA node - there should be no warning
• if system doesn't support NUMA at all (old 32bit systems) there should not be warning too

There's a problem with test sets that it doesn't contain any information about netdev's NUMA membership so the only way to solve this issue is to test on different hardware. It'll be good to improve tests/rss-ladder.tests/ with this data or at least add it for new dataset from host there this feature will be tested.

[abhi2898]

Hi, I would like to work on this issue.

[strizhechenko]

Ok, let me know if you'll have any questions or need some help!

[abhi2898]

Ok sir

[abhi2898]

Can you please elaborate a liitle more on the problem?

[strizhechenko]

Well the problem.

By default this utilities automatically detect NUMA node (you may take a look at topology.py, pci.py and base_tune.py) of network device and choose related CPU socket to bind.

Sometimes people intentionally bind network device to foreign NUMA node because local NUMA node is overloaded by already processing traffic from 3+ network devices. It's usually temporary solution to gain better performance until they replace their hardware or at least put NIC into other PCI slot. But nothing is as long-term as temporary! This workaround may be forgotten and it's really bad. So, I want to add an warning-reminder for that case.

[abhi2898]

Do I need to specify the process, running on the foreign node or just I need to raise an alarm that some process is running on foriegn node.

[strizhechenko]

You need to raise an alarm in case of binding network device queue to foreign node, yeah. (It's just warning, not fatal error).

[abhi2898]

Hi, I am on verge of completing this task, will take some more time to submit the final code.

[strizhechenko]

Hi again! It's nice, BTW you can submit intermediate code.