accel-ppp — Police Upstream, TBF Downstream
accel-ppp applies one enforcement engine per direction, and the choice falls out of how Linux handles each path:
- Upstream is ingress at the BNG, where queuing is not available inline — so accel-ppp meters and drops with the clsact policer. Cheap, near-zero added latency.
- Downstream is egress toward the subscriber, where a queuing qdisc attaches naturally — so accel-ppp shapes with TBF, smoothing bursts toward the customer instead of dropping them.
Both directions support distinct rates and bursts, dynamic provisioning via RADIUS, mid-session changes via CoA, and built-in time-of-day rate tiers. Neither direction performs color marking.
3. Per-Session Functionality
Table 1 compares what each system can do at the level of an individual subscriber session.
Table 1 — Functionality per session
| Capability | Quantum (tri-color, per-direction) | accel-ppp (police up / tbf down) |
|---|---|---|
| Metering model | 3-color (srTCM / trTCM) | 2-color (conform / exceed) |
| Rates per direction | CIR + second rate (EBS or PIR/PBS) | Single rate |
| Independent params per direction | Yes (distinct CIR/PIR/burst each way) | Yes (distinct rate/burst each way) |
| Burst structure | Two buckets (CBS + EBS/PBS) | One bucket (burst-factor + cburst) |
| Outcomes | Green / Yellow / Red | Conform / Exceed |
| Upstream action on exceed | Mark (yellow) or drop (red) | Drop (policer) |
| Downstream action on exceed | Mark (yellow) or drop (red) | Queue + delay (shaper) |
| DSCP / drop-precedence marking | Yes, from meter result | No |
| Color-aware mode | Typically yes | No |
| Marking propagates to core (differential WRED) | Yes | No (edge-only) |
| Burstable-above-committed tier | Yes (yellow admitted) | No (single committed rate) |
| Per-session granularity | Yes | Yes |
| Dynamic rate via RADIUS | Yes | Yes (Filter-Id, attr-up / attr-down) |
| Mid-session rate change (CoA) | Yes | Yes |
| Time-of-day rate tiers | Vendor-dependent (often no) | Yes (built-in time-range) |
| Selective bypass of metering | Vendor-dependent | Yes (fwmark) |
4. Performance and Suitability
Table 2 assesses each system against the outcomes an operator actually cares about. "Winner" reflects the better fit for that single criterion; it is not an overall score.
Table 2 — Performance and suitability
| Criterion | Winner | Why |
|---|---|---|
| Best consumer experience — single flat tier | ~Tie, edge to accel-ppp | TBF shapes downstream (smooths bursts, fewer drops to the customer); Quantum would drop/mark red. For one flat speed, shaping feels cleaner. |
| Best consumer experience — premium / burstable tiers | Quantum | Yellow traffic is admitted-and-marked, so "committed X, burst to Y" survives unless the core is congested. accel-ppp has no burstable tier. |
| Lowest latency — upstream | ~Tie | Both meter inline without queuing. Quantum may mark-admit yellow rather than drop, reducing TCP retransmits. |
| Lowest latency — downstream (at the BNG) | Quantum | Tri-color marker does not buffer; TBF adds small, bounded queuing delay by design. |
| Lowest packet loss for bursts — downstream | accel-ppp | Shaping holds bursts briefly instead of dropping; Quantum drops red immediately. |
| Best for network-wide congestion management | Quantum | Per-session coloring propagates into the core for differential drop. accel-ppp enforcement is edge-only. |
| Best for oversubscription / admission control | Quantum | Two-rate metering + marking is purpose-built for oversubscribed committed/peak contracts. |
| Operational simplicity and automation | accel-ppp | RADIUS-driven, CoA, time-of-day tiers, simple config, fewer moving parts. |
| CPU / cost efficiency per session | accel-ppp | police + TBF is lightweight on commodity x86; no marking/DPI pipeline. |
| Horizontal scalability on commodity hardware | accel-ppp | Scales out across inexpensive nodes; open source, no per-session licensing. |
| Flexibility (dynamic + scheduled rates) | accel-ppp | Built-in time ranges + CoA make dynamic / scheduled tiering easy. |
5. Analysis
The comparison splits cleanly along a sophistication-versus-simplicity axis.
Quantum's tri-color QoS wins wherever the value lives in the metering and marking layer: burstable and oversubscribed product tiers, congestion-aware fairness, and getting the core network to cooperate on differential drop. That is genuine network-engineering capability that accel-ppp does not have — the absence of a yellow/marked state and of color propagation is structural, not a tuning gap.
accel-ppp's standard police/TBF configuration wins on cost, operational simplicity, scalability on commodity hardware, and dynamic / time-of-day flexibility. It is also competitive on plain single-tier consumer experience, because downstream shaping smooths bursts rather than dropping them.
The one counterintuitive result is worth stating plainly: on the downstream path accel-ppp can deliver lower packet loss (it buffers) while Quantum delivers lower latency (it does not). "Better" there depends on whether the subscriber base is more sensitive to delay or to loss.
Guidance for selection. If the roadmap includes burstable committed/peak tiers, oversubscription with enforced fairness, or end-to-end QoS where the core honors per-subscriber coloring, the tri-color platform is the stronger fit. If the objective is straightforward per-subscriber speed tiers, dynamic provisioning, and low cost per session on commodity hardware, accel-ppp's standard configuration meets the need with far less complexity.
6. Assumptions and Caveats
- Quantum IPoE behavior is modeled from the standard srTCM / trTCM three-color definition; vendor specifics were not available and should be confirmed.
- accel-ppp is assessed in its shipped default (
up-limiter=police,down-limiter=tbf). Enabling HTB in either direction changes the cost and latency profile and adds rate-plus-ceil borrowing, but still does not add color marking. - Performance verdicts describe the cost structure and behavior of each model. Hardware-specific figures (CPU per Gbps, latency under saturation, memory per session) require bench testing — for accel-ppp, BNG Blaster against the real rate distribution is the standard method.