test: extend BenchmarkScheduling with multi-NodePool variants - #3246

Fixes #N/A

Description

Adds BenchmarkSchedulingMultiNodePool to pkg/controllers/provisioning/scheduling/scheduling_benchmark_test.go. The existing BenchmarkScheduling* family drives Scheduler.Solve against a single NodePool, so per-NodePool costs never appear in the numbers. The new sub-benchmarks cover a (NodePoolCount, PodCount) grid of {5, 10, 20} x {100, 500, 1000} NodePools by pods, giving nine cells.

Cross-NodePool cost lives in a few paths that the single-NodePool benches skip: NewTopology -> buildDomainGroups iterates NodePools when constructing domain groups, per-NodePool GetInstanceTypes fans out during scheduling, and per-NodePool scheduling-template construction runs once per NodePool. A regression in any of those paths lands on the row of the grid that exercises it, so ns/op, B/op, and allocs/op deltas point at the offending path rather than getting averaged into a single number.

A helper setupMultiNodePoolScheduler constructs the fixture: nodePoolCount NodePools with generous CPU and memory limits, sharing the same 100-instance-type list from fake.InstanceTypes(100). Scheduler.Solve runs against makeDiversePods(podCount). No new dependencies. No changes to production code.

How was this change tested?

Ran the new benchmark family locally with a single iteration per sub-bench to confirm every cell executes and reports timings:

go test -tags=test_performance -bench=BenchmarkSchedulingMultiNodePool -benchtime=1x -run=^$ ./pkg/controllers/provisioning/scheduling/...
goos: darwin
goarch: arm64
pkg: sigs.k8s.io/karpenter/pkg/controllers/provisioning/scheduling
cpu: Apple M4 Pro
BenchmarkSchedulingMultiNodePool/5NP_100Pods-14           1      777765375 ns/op
BenchmarkSchedulingMultiNodePool/5NP_500Pods-14           1     5859031625 ns/op
BenchmarkSchedulingMultiNodePool/5NP_1000Pods-14          1     9188544375 ns/op
BenchmarkSchedulingMultiNodePool/10NP_100Pods-14          1      738149500 ns/op
BenchmarkSchedulingMultiNodePool/10NP_500Pods-14          1     3523763250 ns/op
BenchmarkSchedulingMultiNodePool/10NP_1000Pods-14         1     7514327959 ns/op
BenchmarkSchedulingMultiNodePool/20NP_100Pods-14          1      614226125 ns/op
BenchmarkSchedulingMultiNodePool/20NP_500Pods-14          1     4771892084 ns/op
BenchmarkSchedulingMultiNodePool/20NP_1000Pods-14         1     9333993542 ns/op
PASS
ok      sigs.k8s.io/karpenter/pkg/controllers/provisioning/scheduling   49.377s

make verify, make vulncheck, and make licenses all pass locally on the rebased branch. Numbers above are with -benchtime=1x for a smoke-test only; longer runs are needed before treating any single cell as a baseline.

AI Disclosure

I used an LLM to parallel program with me on this. That included drafting the PR description and running the rebase against upstream/main. All code changes on this branch were authored by me previously; the LLM did not write the benchmark code.

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.