On Tue, Nov 09, 2021 at 09:58:34PM +0800, Gang Li wrote:

On 11/9/21 8:12 PM, Mel Gorman wrote:

quoted

That would be a policy decision on how existing tasks should be tuned
if NUMA balancing is enabled at runtime after being disabled at boot
(or some arbitrary time in the past). Introducing the prctl does mean
that there is a semantic change for the runtime enabling/disabling
of NUMA balancing because previously, enabling global balancing affects
existing tasks and with prctl, it affects only future tasks. It could
be handled in the sysctl to some exist

0. Disable for all but prctl specifications
1. Enable for all tasks unless disabled by prctl
2. Ignore all existing tasks, enable for future tasks

While this is more legwork, it makes more sense as an interface than
prctl(PR_NUMA_BALANCING,PR_SET_NUMA_BALANCING,1) failing if global
NUMA balancing is disabled.

Why prctl(PR_NUMA_BALANCING,PR_SET_NUMA_BALANCING,1) must work while global
numa_balancing is disabled? No offense, I think that is a bit redundant.

For symmetry and consistency of the tuning. Either there is per-process
control or there is not. Right now, there is only the ability to turn
off NUMA balancing via prctl if globally enabled. There is no option to
turn NUMA balancing on for a single task if globally disabled.

And
it's complicated to implement.

That is true.

It's hard for me to understand the whole vision of your idea. I'm very
sorry. Can you explain your full thoughts more specifically?

I'm not sure how I can be more clear.

----------------------------------------------------

Also in case of misunderstanding, let me re-explain my patch using circuit
diagram.

I understood what you are proposing. In your case, global disabling
is an absolute -- it's disabled regardless of prctl therefore
prctl(PR_NUMA_BALANCING,PR_SET_NUMA_BALANCING,1) has no meaning and it
either does nothing at all or fails so why does the option even exist?

Why global numa_balancing has high priority? There are two reasons:
1. numa_balancing is useful to most processes, so there is no need to
consider how to enable numa_balancing for a few processes while disabling it
globally.
2. It is easy to implement. The more we think, the more complex the code
becomes.

Of those two, I agree with the second one, it would be tricky to implement
but the first one is less clear. This is based on an assumption. If prctl
exists to enable/disable NUMA baalancing, it's possible that someone
else would want to control NUMA balancing on a cgroup basis instead of
globally which would run into the same type of concerns -- different
semantics depending on the global tunable.

-- 
Mel Gorman
SUSE Labs