Machine Types¶

Google Cloud offers different machine types that vary in CPU architecture, memory, and price. Cloud Batch provisions VMs of the chosen machine type to run each pipeline task.

For how cpuMilli and memoryMib map to machine resources, see Cloud Batch: Resource units. For recommended configurations per stage, see Sizing Recommendations.

Machine family overview¶

A machine family groups machine types that share the same CPU platform and pricing model. The pipeline primarily uses C4D (AMD compute-optimized) for its strong single-thread performance at a competitive price, but other families are available depending on your workload.

Here are some examples of machine families that can fit your needs:

For details on each machine families and pricing, see:

Machine families resource and comparison guide | Google Clouddocs.cloud.google.com

VM instance pricing | Google Cloudcloud.google.com

Machine type selection¶

How it works¶

When Cloud Batch runs a job, it provisions VMs to execute tasks. The workflow accepts an ordered candidate chain per stage. If a candidate cannot obtain enough capacity, the workflow cancels and drains that Batch job before trying the next candidate. A task that fails with a container exit code (for example 1, or 137 for out-of-memory) stops the workflow with TASK_FAILED instead, since another machine would fail the same way; only Batch VM errors (reserved 500xx codes) and capacity stalls fall back.

Automatic fallback: Configure machine_types in preferred order. Every candidate is validated against the stage's CPU and memory minimums before submission.

Explicit CLI pin: Pass a --stage-*-machine-type option to use exactly one machine type and disable fallback for that stage. This is useful for controlled benchmarks and debugging.

For details on how Batch automatically creates and deltes resources that meet specification, see:

Job resources | Cloud Batch | Google Clouddocs.cloud.google.com

Pipeline configuration¶

In the pipeline, machine type selection is configured per stage:

google_cloud:
  batch:
    stage_b:
      cpu_milli: 2000
      memory_mib: 7168
      machine_types:
        - c4d-standard-2
        - c4-standard-2
        - n4d-standard-2
        - n4-standard-2

epycloud run workflow --exp-id EXAMPLE \
  --stage-b-machine-type c4d-standard-2

See Configuration Variables for all per-stage resource settings.

Naming convention¶

Google Cloud machine types follow the pattern: {family}-{type}-{vcpus}

• Family: e2, n2, c4, c4d, etc.
• Type: standard (balanced), highmem (more memory per vCPU), highcpu (more vCPUs per memory)
• vCPUs: Number of virtual CPUs (2, 4, 8, 16, ...)

Examples:

• c4d-standard-2: C4D family, standard memory, 2 vCPUs
• c4d-highmem-4: C4D family, high memory, 4 vCPUs
• e2-standard-4: E2 family, standard memory, 4 vCPUs

For full list of machine types, see:

VM instance pricing | Google Cloudcloud.google.com

Common machine types¶

Machine types commonly used by the pipeline. A typical calibration task in Stage B for a single state uses approximately 4 GB of memory.

A vCPU is one hardware thread, and two vCPUs share each physical core. A single-threaded task such as Stage B calibration needs 2 vCPUs (one physical core) to run at full speed; two tasks on one core each run about 1.9x slower. See CPU per task and tasks per VM. For per-stage defaults and guidance, see Sizing Recommendations.

2-vCPU machine types¶

4-vCPU machine types¶

Use these for memory-intensive workloads such as Stage C (output aggregation), which loads all Stage B results into memory.

For Cloud Build machine types, see Cloud Build.