Three tools, three different trade-offs
A recurring survey of the field frames the choice clearly: network operators and researchers constantly need a platform to validate a proposed system, protocol, or security control, and the recurring challenge is balancing traffic realism, scalability, and cost. Simulators and emulators offer a controlled, reproducible environment but sacrifice some realism in the process; testbeds offer realism and scale, at the cost of being slower to set up and more expensive to run.
Simulators
Simulators represent network components with numerical and behavioral models, advancing the simulation state through discrete events — a packet arrival, a timer firing — rather than modeling continuous physical processes. Because everything runs as software on ordinary hardware, results are highly reproducible and experiments are relatively cheap to run at scale. The trade-off is fidelity: a simulator’s output is only as accurate as the models feeding it, and industrial field reports on network simulation note that simulation aims to mimic a system’s internal states, which can require significant modeling effort to get right, especially for physical-layer effects.
Emulators
Emulators sit between simulators and physical testbeds. Rather than modeling a component’s internal state, an emulator aims to reproduce its external behavior, and critically, it can interact with real network traffic and real software — a real TCP/IP stack, or real application code, running on top of an emulated network path. Early emulation tools such as NIST Net demonstrated this directly, letting experimenters impose realistic delay, jitter, bandwidth limits, and packet loss on live IP packets, combining much of a simulation’s controllability with much of a live network’s realism. The cost is usually reduced scale compared to a pure simulator, since emulation typically consumes real compute and network resources roughly in proportion to the traffic being emulated.
Testbeds
Testbeds range from small physical lab setups to large federated platforms such as GENI, and they trade the previous two approaches’ convenience for realism: real hardware, real operating systems, and real network stacks under test, with performance characteristics that do not need to be modeled because they are simply measured. That realism comes at real cost — in equipment, physical space, and the time needed to reconfigure a topology — which is precisely why simulators and emulators remain the default first step in most research workflows, with testbed validation reserved for later stages.
A practical decision guide
| Question | Points toward |
|---|---|
| Do you need to sweep hundreds of parameter combinations quickly? | Simulator |
| Do you need real application or OS code exercised against controlled network conditions? | Emulator |
| Do you need to demonstrate behavior on real hardware for publication or deployment sign-off? | Testbed |
| Is budget or lab space the binding constraint? | Simulator or emulator |
| Is the research question specifically about simulation or emulation fidelity itself? | Testbed, as ground truth for validation |
They are usually not mutually exclusive
In practice, the strongest research designs use more than one of these tools in sequence: a simulator or emulator to explore a wide parameter space cheaply, followed by a smaller-scale testbed run to confirm the most promising configurations behave the same way on real hardware. That pairing is exactly the methodology question NSTools was founded to examine — the workshop’s first founding theme, the evaluation of simulation accuracy, is essentially the study of how well each of these tools’ predictions hold up against the next one in the chain.
A note on terminology
Not every paper uses these three terms the same way — some use “simulation” loosely to cover tools that are technically emulators. When reading or reviewing a paper, it is worth checking explicitly whether real traffic or real code is involved anywhere in the evaluation, rather than relying on the label the authors chose.