Build an Accelerator Quota Allocation and Over-Subscription Planner

Model multi-team GPU reservation quotas against historical compute usage traces to identify over-subscription capacity. Generate team quota adjustment proposals with hosted Claude without risking OOM evictions.

Compute Infrastructure Manager

Ingest team allocation quotas alongside historical 30-day GPU time-slice usage logs.

Calculate concurrency peaks, idle time-blocks, and safe over-subscription ratios.

Use hosted Claude to draft quota restructuring recommendations and SLA impact analyses.

Publish an over-subscription plan with scheduled reclamation guardrails for cluster admins.

Quota planning portal featuring historical demand duration curves and adjustable over-subscription sliders.

Construct synthetic JSON traces showing hourly GPU reservation claims versus actual compute activity per team. Starter examples below are synthetic. They are not downloaded records.

NVIDIA DCGM exporter Applies GPU compute and memory utilization metrics as defined in standard DCGM monitoring systems.

optuna, anthropic-claude, pydantic-ai. optuna optimizes numerical over-subscription ratios to minimize unused capacity while capping eviction probability. Hosted Claude API for narrative justification; local numerical optimization via Optuna.

Submit runnable source, pinned dependencies, setup commands, a working interface, example outputs, and a recorded end-to-end demo. Include tests and an action log.

Calculate the risk score of queue starvation under simulated over-subscription levels.

Synthesize team-by-team quota adjustment cards with transparent historical rationale.

Enforce strict priority rules preserving critical training workloads over batch eval jobs.

```json [ { "input": { "cluster_total_gpus": 64, "allocations": [ { "team": "ad-ranking", "assigned_gpus": 32, "peak_concurrent_gpus": 18, "avg_utilization_pct": 45 }, { "team": "speech-ai", "assigned_gpus": 32, "peak_concurrent_gpus": 20, "avg_utilization_pct": 50 } ] }, "expected_output": { "current_effective_headroom_gpus": 26, "recommended_over_subscription_ratio": 1.35, "proposed_reallocations": [ { "team": "ad-ranking", "new_guaranteed_gpus": 20, "burst_gpus": 12 }, { "team": "speech-ai", "new_guaranteed_gpus": 22, "burst_gpus": 10 } ], "unlocked_free_pool_gpus": 22 } }, { "input": { "cluster_total_gpus": 16, "allocations": [ { "team": "core-nlp", "assigned_gpus": 16, "peak_concurrent_gpus": 16, "avg_utilization_pct": 95 } ] }, "expected_output": { "current_effective_headroom_gpus": 0, "recommended_over_subscription_ratio": 1, "proposed_reallocations": [ { "team": "core-nlp", "new_guaranteed_gpus": 16, "burst_gpus": 0 } ], "unlocked_free_pool_gpus": 0 } } ] ```

Analytical quota modeling tool only; does not modify scheduler cgroups or evict running jobs. Use mock external writes. Require human review before any real deployment.