PROFILE — Cluster Profile Engine
PROFILE is the first engine in every PTL assessment. It characterizes your cluster — scheduler type, GPU fleet, telemetry sources, MIG configuration — and produces the routing manifest that all downstream engines use. PROFILE is not scored. It is a prerequisite.
What PROFILE does
Section titled “What PROFILE does”Scheduler detection. PROFILE identifies your scheduler type — Slurm, Kubernetes, or Ray — and records the input path each downstream engine should use.
GPU fleet characterization. PROFILE maps your hardware inventory to PTL model keys. CORE uses this to compute hardware-workload fit. PROFILE also detects MIG (Multi-Instance GPU) configuration when present.
Telemetry source routing. ACE supports four input paths: CLAW telemetry, DCGM metrics, Kubernetes pod metrics, and Slurm sacct. PROFILE evaluates which sources are available and sets the priority order — CLAW takes precedence, then DCGM, then Kubernetes metrics, then sacct.
Output
Section titled “Output”PROFILE produces a routing manifest JSON with:
scheduler_type— detected scheduler (slurm, kubernetes, ray)ace_input_path— preferred telemetry source for ACEpace_input_path— scheduler path for PACEgpu_fleet— list of GPU model keys with countsmig_enabled— boolean; whether any GPUs are MIG-partitionedprofile_version— engine version string
GRADE and ATLAS receive the profile path and use it to include a cluster profile section in the certification report.
Data requirements
Section titled “Data requirements”A standard PROFILE run requires hardware inventory — GPU model and count per node. If running in Slurm mode, sacct output provides job history for ACE and PACE. CLAW automates profile collection for organizations where it can be deployed.