Add custom image support

[jasminetMSFT]

Summary

Enables the use of custom container images in image pull benchmarks, allowing more realistic testing with large, multi-layer images that match real customer workloads.

Changes

• Custom image support in deployment templates: Added TestImage parameter to deployment_template.yaml to allow overriding the default test image. TestImage is intentionally scoped to Linux/memory/AKS, this is the only path used by image-pull scenarios where we need custom image sizes for throughput testing. CPU workloads and Windows use resource-consumer which has specific stress commands.
• CRI module updates: Extended cri.py and execute.yml to pass through the test_image parameter from pipeline configuration
• Updated image-pull-n10 pipeline:
  • Configured to use benchmark/customer-replica:v1 (10GB, 79-layer image matching customer manifest)
  • Increased timeouts: operation_timeout: 30m, pod_startup_latency_threshold: 600s
  • Reduced max_pods from 30 to 26 to prevent pod scheduling failures due to node resource constraints

Testing

Validated with 10-node cluster pulling 10GB custom image:

• Pipeline: link

[Copilot]

Pull request overview

Adds support for overriding the container image used by the ClusterLoader2 CRI/image-pull benchmark so pipelines can exercise large, multi-layer “customer-like” images.

Changes:

• Plumbs a new test_image/CL2_TEST_IMAGE parameter from the pipeline step into the CRI override generator (cri.py) and CL2 config.
• Updates the deployment template to use a configurable image (AKS/Linux/memory path) and adds a topology spread constraint to improve node distribution.
• Refines the containerd throughput “AvgPerNode” PromQL query to exclude nodes with no pulls.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
steps/engine/clusterloader2/cri/execute.yml	Passes --test_image into the CRI override step via env/config.
modules/python/clusterloader2/cri/cri.py	Adds --test_image CLI arg and writes CL2_TEST_IMAGE into overrides.
modules/python/clusterloader2/cri/config/deployment_template.yaml	Introduces TestImage template param, uses it for the AKS/Linux/memory image, and adds topology spread constraints.
modules/python/clusterloader2/cri/config/containerd-measurements.yaml	Filters “AvgPerNode” to active nodes in PromQL.
modules/python/clusterloader2/cri/config/config.yaml	Wires CL2_TEST_IMAGE into the deployment template fill map.

Comments suppressed due to low confidence (1)

modules/python/clusterloader2/cri/config/deployment_template.yaml:59

• When TestImage is not the default, this template no longer renders an explicit command/args for the memory container (the stress block is gated by if eq $TestImage ...). That changes the benchmark from a known long-running workload to whatever the image entrypoint does, which can cause early exits or add non-pull-related variance. Consider keeping a stable long-running command for all images, or making the command configurable alongside TestImage.

          {{if eq $TestImage "e2e-test-images/resource-consumer:1.13"}}
        command:
          - stress
        args:
          - --vm

[liyu-ma]

LGTM.