ncsim is a lightweight discrete-event simulator for networked computation. It lets you place DAG tasks on heterogeneous nodes, route the resulting data flows over multi-hop networks, and see what happens when those wireless transfers actually contend and interfere.
Illustrative 4×4-grid case from our experiments: HEFT 14.9 s vs. round-robin 25.7 s without interference; under CSMA/CA, HEFT 1244 s vs. round-robin 523 s.
The visualization shows the wireless topology, active transfers, Gantt-style execution timeline, and event log together, so you can see how a placement decision turns into network contention over time.
Traditional DAG schedulers often treat communication time as data size ÷ a fixed pairwise bandwidth. On a wireless mesh, the scheduler’s placement changes which routes are needed; those routes determine which links are active; and those links change each other’s effective rates.
Assigning dependent tasks to different nodes creates transfers. A placement that looks compute-efficient can generate a large communication footprint.
Multi-hop paths activate multiple relays and contention domains. The widest path for one flow may be a poor choice once many flows run together.
As flows start and finish, effective link rates change. Transfer completion times move, task readiness moves, and the realized makespan can diverge sharply from the planner’s estimate.
ncsim sits between packet-level network simulators and workflow simulators: more wireless-realistic than fixed-bandwidth DAG tools, but much lighter than full packet simulation when you want to sweep schedulers, routes, topologies, workloads, and RF parameters.
Pluggable scheduler, routing, and interference components; dynamic bandwidth recalculation as transfers start and finish.
In a 108-run factorial study, the scheduler selected by an interference-free evaluation changed once realistic wireless interference was included in 5 of 18 scheduler-comparison cases.
A locality-biased HEFT model that strongly discourages non-adjacent placements dramatically outperformed a widest-path HEFT model across the configurations studied.
ncsim’s analytical/flow-level Wi-Fi model was validated internally, against Bianchi’s published results, and against packet-level ns-3 experiments.
Single-flow “best path” logic does not necessarily survive shared wireless airtime. In our studies, shorter routes often beat wider ones because they activate fewer relays and contention domains.
Test HEFT, CPOP, PEFT, Min-Min, Sufferage, round-robin, or your own scheduler under the same executable network model.
Ask whether a placement remains good after its data dependencies are mapped onto real multi-hop paths.
Prototype new schedulers or routing policies that use locality, contention state, or congestion feedback.
Vary topology, density, compute heterogeneity, payload size, Wi-Fi parameters, and concurrent workflows.
Use deterministic event traces and metrics for analyses, plots, algorithm debugging, and artifact evaluation.
Show students why “communication cost” is not just a fixed number in distributed wireless systems.
Macroscopic scheduler evaluation, bulk DAG-edge transfers, static or slowly varying deployments, contention-driven Wi-Fi effects, parameter sweeps.
You need packet/protocol engineering, detailed retransmission behavior, fast mobility/fading, TCP slow-start effects for short flows, or fine PHY dynamics.
ncsim intentionally keeps first-order wireless effects that can change scheduling outcomes—distance-dependent PHY rate, CSMA/CA contention, hidden terminals, link sharing—while avoiding packet-level detail that would make large scheduler sweeps cumbersome.
# Clone the full repo
git clone https://github.com/ANRGUSC/ncsim.git
cd ncsim
pip install -e .
# Run a demo
ncsim --scenario scenarios/demo_simple.yaml \
--output results/demo
# Inspect the event trace
python analyze_trace.py results/demo/trace.jsonl \
--gantt --timeline --tasksPython 3.12+. You can also install the core package with pip install anrg-ncsim.
The repository includes a CLI and web visualization. The UI can configure scenarios and replay simulations with synchronized network activity, task Gantt charts, and an event log.
We would love to see researchers use it for new scheduling, routing, dispersed-computing, edge-AI, and tactical-networking questions—and to contribute models, scenarios, and algorithms back to the project.