Skip to content

Add multi-variants benchmark for testing WVA bejaviour #1425

Description

@biranofer

Component

Setup/Standup

Desired use case or feature

A major value aspect of WVA is in multi-variants scenarios (there the model-scope optimization is a differentiator compare to KEDA/HPA based auto-scaling). We need to have the ability to run the existing scenarios (as prefill_heavy) and more in multi-variants setup (can start with two-variants). Will need a simple way of defining the variants (hardware, parallelism, cost etc.) run the benchmarks and canalize the results (additional per-variant graphs etc,).

Proposed solution

An experimental benchmark setup and run is described here
https://github.com/biranofer/workload-variant-autoscaler/blob/feat/two-variant-wva-benchmark/docs/developer-guide/two-variant-wva-benchmark.md, here is a resulted graph from a cost-differentiated two variants (primary and v2) run, using the prefill_heavy scenario:

Image

The example implementation matching the design above is in the two-variant-wva-benchmark branch at my fork

Alternatives

No response

Additional context or screenshots

No response

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions