All projects

GPUShare

An OMNeT++ simulation of pooled GPU leasing across two campus network segments, with capacity beacons, centralized scheduling, background traffic, and repeatable analysis.

Status
Course project · reproducible simulation
Role
Designed and implemented the OMNeT++ network/control-plane modules, experiment configurations, recorded statistics, and Python analysis workflow.
Networks · Discrete-event simulation2026
  1. 01Job client
  2. 02Network bus
  3. 03Scheduler
  4. 04GPU host
High-level flow only. Open the case study for decisions, evidence, and limitations.

Campus GPU capacity is unevenly distributed. The project models how clients, shared links, a router, a scheduler, and heterogeneous hosts interact when GPU jobs compete with burst traffic.

  • Implemented the bus, router, GPU host, scheduler, client, and background-flow modules without the INET framework.
  • Recorded JCT, utilization, queue length, routed traffic, and per-bus volume across seeded runs.
  • Documented scope honestly: simulation results are not general performance guarantees.

A complete modeled path

Clients submit jobs, a central scheduler selects a host from capacity beacons, the host reserves a slot, and completion is observed end to end across the modeled network.

Measured, not generalized

Committed scenarios compare no, light, medium, and heavy burst traffic. The results belong to this topology, workload, configuration, and seed set.