scx_lavd: Use Peek operation in lavd if available

[rrnewton]

See commit description for the first diff, copied below.

---

[SCX][lavd] modify lavd to use new peek operation
Summary:

The per-cpu queues mode in LAVD checks multiple queues for potential steal targets.
These checks can be made more efficient with the forthcoming O(1) lockless peek operation.
This patch updates LAVD to use peek in three places:

• try_to_steal: picking cpu with lowest vtime
• on consume_task: checking the cpu DSQ
• on consume_task: checking the domain DSQ

Note that these usages are necessarily racy because no lock is held. That is, we
could peek a low vtime, but the task may have moved or completed by the time we
attempt to steal it.

Test Plan:

• ran schbench under a VM [1 cpdom, 50cpus]

Reviewers: daidavid

[rrnewton]

@daidavid - I'll post some more baremetal perf results here. Under a VM with 50 cpus, 1 cpudom, 120 GB ram, it really doesn't move schbench score.

[daidavid]

@rrnewton Yes, let's get the numbers with baremetal. VM numbers are noisy and not very meaningful, a few things to keep in mind:

• I wouldn't expect the new load balancing to help with schbench that much since vtime ordering isn't that important for a microbenchmark. I'd test it on some real workloads if you can.
• I have doubts about how often the task steal path actually gets used given all of the recent changes, it would be interesting to instrument that. We might need to update the enqueue logic as well.

[multics69]

Thank you, Ryan! I left a few comments for the next version.

One nitpicking is to use "scx_lavd" instead of "[SCX][LAVD]" as a prefix of the commit message for consistency.

[rrnewton]

@multics69 - ok, comments resolved. Let me know if that looks good now.

[multics69]

@multics69 - ok, comments resolved. Let me know if that looks good now.

Thanks for addressing the comments. It looks good now. Let's merge it to the main once the kernel patch is submitted.