Tuning Profiles

The enterprise runtime exposes three tuning profiles. A profile is a request, not a guarantee. Use doctor to verify what the host actually honored.

Profiles

ProfilePostureTypical UseHost Expectations
defaultBackoff-oriented scheduling, no privileged assumptionsLocal development, smoke tests, broad compatibilityNormal Linux host
low-latencySpin wait, balanced pinning intent, mlockall requestLower-latency hosts where privilege is availableMemory-lock support, sufficient CPU headroom
isolated-coreSpin wait, isolated CPU intent, RT scheduling, locked memoryProduction posture for tightest latency and determinismIsolated CPU set, RT scheduling, memory-lock, correct SHM placement

Profile Behavior

The runtime distinguishes between four states for each tuning request:

StateMeaning
RequestedWhat the profile asked for
AppliedWhat the host successfully applied
DegradedWhat was requested but could not be fully applied
FailedHost-readiness failures that prevented application

This split matters for operations. A malformed request and an under-provisioned host are different problems and should not be debugged the same way.

The doctor Verb

Always validate the requested posture on the target host with doctor before treating it as production-ready.

doctor checks:

  • Memory-lock availability and limits
  • CPU isolation and pinning capabilities
  • RT scheduling permissions
  • SHM placement and permissions
  • Available CPU headroom

Effective Tuning in Cloud

Cloud deployments surface the effective tuning record in the deployment details view:

  • Requested settings from the config bundle
  • Applied settings reported by the agent
  • Degraded settings with reasons
  • Failure reasons
  • Host observations

This gives operators a single view of what was asked for and what actually happened on the target host.

Operational Guidance

  • Start with default during development and testing
  • Use low-latency when the host has privilege but full isolation is not guaranteed
  • Reserve isolated-core for production hosts that pass doctor validation
  • Always run doctor on new hosts before assuming production readiness
  • Monitor the degraded/failed split in Cloud deployment details to catch host configuration drift
Provenance
Need the canonical source?
Use the public hub to orient yourself, then jump to repo-owned docs or rustdoc when you need contract-level detail.