Personal project · job-scheduler

Topology-aware workload management

An inspectable workload-management model with per-cluster queues, topology-aware resource placement, transactional allocation, and a React interface for cluster and job state.

Actual job-scheduler interface showing a 3D torus cluster and node resource details
Screenshot from the job-scheduler repository

The question

How should queue order, resource availability, and hardware topology influence an allocation? This project makes those decisions visible without requiring a physical cluster.

What is implemented

FIFO and priority queues, deterministic Pack and Spread placement, multi-node allocations, institute quotas, reservations, lifecycle history, and an interactive React cluster view. PostgreSQL stores individual resource units and allocation mappings.

The architectural boundary

The scheduler selects the next placement attempt. A transaction-local placer chooses specific resource units from locked inventory. The server coordinates policy and persistence; API routes own commit and rollback.

Network topology and GPU discovery

The leaf–spine model registers existing compute Nodes through their NIC endpoints and connects them to switch ports and physical links. Path discovery checks operational state, configured capacity, and latency; GPU candidate queries combine those paths with the existing resource-unit inventory.

Engineering notes
The actual application

Cluster state and job history.

Actual cluster overview listing ring, torus, and mesh clusters
Cluster overview · repository screenshot
Actual job page with resource requirements, allocation, and lifecycle events
Job detail · repository screenshot
Inside the project

Classes & relationships.

Open a diagram to inspect the full-size SVG ↗
Runtime responsibilities

Class architecture

Open full size ↗
Server, Scheduler, Placer, their policies, and the topology view. Redrawn from the implemented class diagram in the project documentation.
Server, Scheduler, Placer, their policies, and the topology view. Redrawn from the implemented class diagram in the project documentation. Project source ↗
Persisted state

Entity relationships

Open full size ↗
The core ownership, request, allocation, and hardware tables, rendered from actual ORM fields and foreign keys. Open the full schema for quotas, reservations, and usage-table relationships.
The core ownership, request, allocation, and hardware tables, rendered from actual ORM fields and foreign keys. Open the full schema for quotas, reservations, and usage-table relationships. Project source ↗
The project’s original ERD drawings

The original whiteboard and subsystem SVGs are preserved here as design history. The live schema has evolved; the current-model ERD above follows the ORM rather than the older lookup-table design.

Original whiteboard ERD from the job-scheduler repository, showing the historical schema design