Identify the important dimensions of a space¶

When working with a high-dimensional configuration space it is natural to ask which dimensions have the greatest influence on a specific experimental outcome. For instance, if a workload has 20 tunable parameters you might want to identify which ones most significantly affect throughput. Understanding this can help narrow future explorations to only the most impactful parameters, reducing the time and resources spent by ignoring those that are less relevant.

The workloads in the ml_multi_cloud dataset are defined by four parameters — provider, cpu_family, vcpu_size, and nodes — giving the entity space four dimensions. This walkthrough finds which of those dimensions most influence wallClockRuntime.

We will:

1. Install the ray_tune operator
2. Reuse the discovery space from the random walk example
3. Run a Latin Hypercube sampling operation with an InformationGain stopper
4. Interpret the ranked dimension output

git clone https://github.com/ibm/ado.git
cd ado

ado create operation -f examples/ml-multi-cloud/lhc_sampler.yaml --set "spaces[0]=$DISCOVERY_SPACE_IDENTIFIER"

Step 1 — Install the ray_tune operator¶

The ray_tune operator is distributed as a separate package:

pip install ado-ray-tune

Confirm it is registered:

ado get operators

You should see ray_tune listed:

Available operators by type:
┌───────┬─────────────┬─────────┬─────────┐
│ INDEX │ OPERATOR    │ VERSION │ TYPE    │
├───────┼─────────────┼─────────┼─────────┤
│ 0     │ random_walk │ 2.0.0   │ explore │
│ 1     │ ray_tune    │ 2.0.3   │ explore │
│ 2     │ rifferla    │ 2.0.3   │ modify  │
└───────┴─────────────┴─────────┴─────────┘

Step 2 — Reuse the discovery space¶

This example uses the same discovery space created in the random walk example. Retrieve its identifier:

ado get spaces

Note the identifier — you will pass it to the operation in the next step.

Step 3 — Run a Latin Hypercube sampling operation¶

The file examples/ml-multi-cloud/lhc_sampler.yaml configures ray_tune to sample using the Latin Hypercube Sampler and stop automatically when the most influential dimensions are known:

ado create operation -f examples/ml-multi-cloud/lhc_sampler.yaml --set "spaces[0]=$DISCOVERY_SPACE_IDENTIFIER"

The operation is configured to:

• Sample up to 32 points using the Latin Hypercube Sampler — a method that maintains properties similar to true random sampling while keeping samples more evenly spread across the space.
• Monitor the mutual information between each dimension and the target metric (wallClockRuntime).
• Stop early via the InformationGain stopper once the most influential dimensions are identified.

RayTune logs progress to the terminal as it runs. Because the InformationGain stopper triggers before all 32 samples are collected, the operation finishes early.

Step 4 — Interpret the results¶

Within the logs for the last sample you will see output like:

(tune pid=72306) Stopping criteria reached after 14 samples.
(tune pid=72306) Total search space size is 48, search coverage is 0.2916666666666667.
(tune pid=72306) Entropy of target variable clusters: 0.6517565611726529 nats.
(tune pid=72306) Result:
(tune pid=72306)     dimension  rank        mi  uncertainty%
(tune pid=72306) 3       nodes     1  0.491089      0.753486
(tune pid=72306) 1  cpu_family     2  0.352622      0.541033
(tune pid=72306) 0    provider     3  0.034638      0.053146
(tune pid=72306) 2   vcpu_size     4  0.000929      0.001425
(tune pid=72306)
(tune pid=72306) Pareto selection:['cpu_family', 'nodes']

Reading the ranked dimension table¶

The dimensions are ranked by mutual information (mi) with the target variable wallClockRuntime. The columns are:

In this run nodes contributes ~75% of the explainable variance, making it by far the most important dimension, while vcpu_size is nearly irrelevant.

The Pareto selection¶

At the end of the output the stopper reports a Pareto selection — the smallest set of dimensions whose combined mutual information exceeds a threshold (0.8 by default).

The stopper computes this as follows:

• For each possible set size, it finds the subset that explains the most mutual information. For example, for size 2 it evaluates all 6 pairs: [nodes, vcpu_size], [nodes, provider], [nodes, cpu_family], etc.
• This produces one optimal set per size (1, 2, 3, 4 in this case) — the Pareto-optimal sets.
• The smallest set whose total mutual information exceeds the threshold is selected.

Summary¶

What's next¶