Abstract
This project reworks the switch datapath of P4Sim, the P4 switch module for the ns-3 network simulator. Before this work, packets reached the channel through an extra port layer, switch links reused the generic CSMA/P2P channel models, and the queues were drained by a timer that polled them regardless of port state. The result was an indirect datapath and a scheduler that did not track when a port was actually free to send.
The work addresses this in three parts. First, it simplifies the switch–channel architecture by removing the intermediate port layer. Second, it uses a more realistic Ethernet channel in place of the older CSMA/P2P channels. Third, it replaces the old polling-based scheduler with an event-driven scheduler built on a Virtual Output Queue (VOQ) structure. Together these give P4Sim a cleaner, more direct datapath and a more realistic model of switch queueing.
Goals
- Simplify the switch–channel architecture and make the data path more direct.
- Use a more realistic Ethernet channel for switch links.
- Replace polling-based scheduling with an event-driven scheduler backed by VOQ.
What We Changed
Simplified Switch–Channel Architecture (PR #22)
- Removed the intermediate port layer.
- The switch NetDevice now connects directly to the channel to send and receive packets.
- This gives a cleaner architecture and a more direct data path, with fewer hops between the switch pipeline and the wire.
More Realistic Ethernet Channel (PR #22; earlier channel work in PR #20 and PR #21)
- Replaced the separate CSMA/P2P channels with a single Ethernet channel.
- This better matches how Ethernet switches are actually connected, so the model behaves more like real switch-to-switch and switch-to-host links.
Event-Driven Scheduling with VOQ (PR #25, PR #26, PR #27, PR #28)
- Replaced the polling-based scheduler with a port-status-driven one.
- Scheduling now runs when an output port becomes idle (signalled by the port finishing a transmission), instead of polling the queues on a timer. This avoids wasted timer wake-ups and matches how hardware reacts to port availability.
- The queues use a Virtual Output Queue (VOQ) structure — traffic is separated per output port — which is more realistic, scales better, and keeps congestion on one port from blocking traffic bound for another.
Links
All code from this GSoC project lives in the following repository: https://github.com/HapCommSys/p4sim
Architecture
Before
In the old design, the switch reached the channel through an intermediate port layer, links used the separate CSMA/P2P channels, and scheduling was polling-based — the scheduler checked the queues on a timer instead of reacting to port state.
After
In the new design, the port layer is gone and the switch NetDevice connects directly to the channel. Links use a single Ethernet channel, and scheduling is event-driven: it runs when an output port becomes idle and is backed by a VOQ structure. The switch–channel refactor landed in PR #22, and the event-driven VOQ scheduling in PR #25–PR #28.
Examples and Tests
The new datapath is checked by a unit suite and a set of example scenarios.
| Scenario | Metric | Result |
|---|---|---|
| Unit suite | test cases | 9 / 9 PASS |
| Conservation / parity | rxBytes / tmReceived / VoqEnqueued / Transmitted / tmDropped | 296000 / 298 / 298 / 298 / 0 |
| Throughput benchmark | achieved line rate (100M / 1G) | 97.13% / 97.09% |
| Strict-priority QoS | HIGH-priority flow kept under load | 95.92% |
| DDoS mitigation | legitimate flow kept (without isolation → with isolation) | 69.86% → 96.04% |
- Unit suite — covers enqueue/dequeue, VOQ behaviour, and event-driven scheduling, 9/9 passing (PR #25, PR #26).
- Conservation / parity (
p4-voq-fabric-integration) — sends a known byte count through the switch and checks the ingress and egress totals match, with no unexplained loss (tmDropped = 0) (PR #28). - Throughput benchmark (
p4-voq-fabric-throughput) — drives the datapath near line rate and reports the achieved rate at 100M and 1G (PR #29). - Strict-priority QoS (
p4-voq-fabric-priority) — a HIGH and a LOW flow share one egress port; the HIGH flow is protected while the LOW flow is throttled (PR #29). - DDoS mitigation (
p4-voq-fabric-ddos) — one legitimate flow against a flood of attacker flows, with a--mitigateflag that turns on per-priority buffer isolation. Without isolation the legitimate flow keeps ~70%; with isolation it keeps ~96% (branchexamples/ddos-mitigation).
To run an example:
./ns3 run contrib/p4sim/examples/p4-voq-fabric-throughput
Future Work
- The Ethernet channel currently models switch links at the link level (data rate and propagation delay); adding configurable error/loss models would broaden the range of conditions P4Sim can reproduce.
- The event-driven scheduler currently supports strict-priority scheduling; adding further policies (for example weighted fair queueing) would let P4Sim model a wider set of QoS behaviours.
- The VOQ datapath is exercised through the V1model architecture; extending it to the PSA and PNA pipelines is a natural follow-up.
References
- ns-3 network simulator — https://www.nsnam.org/
- ns-3 model library (CSMA and point-to-point channel models) — https://www.nsnam.org/docs/models/html/
- P4 language consortium — https://p4.org/
- P4 behavioral model (bmv2) — https://github.com/p4lang/behavioral-model
- P4Sim project repository — https://github.com/HapCommSys/p4sim
- DDoS-mitigation example branch — https://github.com/Vineet1101/P4Simulator/tree/examples/ddos-mitigation