Dropless Routing in Deterministic Networks

• Dropless routing is a network technique that ensures zero packet loss by using proactive data-plane rerouting, multipath diversity, and source-embedded alternate paths.
• It relies on rigorous mathematical formulations like the Delay-Range Constrained Routing (DRCR) problem to optimize cost and delay while enforcing policy compliance.
• Empirical evaluations of methods such as Pulse⁺, D2R, and SlickPackets reveal significantly reduced failover times and minimal packet loss compared to traditional routing protocols.

Dropless routing denotes a class of techniques and architectures that guarantee—within clearly defined system assumptions—zero packet loss (or loss bounded to in-flight packets at microsecond scale) during network failures, reconfiguration, or path changes. The dropless property is fundamental for deterministic networking (DetNet), high-availability fabrics, and ultra-reliable low-latency use cases, and is achieved through fully data-plane rerouting, engineered multipath diversity (1+1 protection), or source-embedded alternate paths. Unlike classical routing methods, which drop packets during control-plane convergence or rely on pre-installed backup state, dropless systems ensure that packets are always forwarded to their destination along a policy-compliant, feasible path, provided connectivity remains.

1. Mathematical Foundations and Problem Formulations

At the core of dropless routing for DetNet lies the Delay-Range Constrained Routing (DRCR) problem, formulated as follows: For a network $G = (V,E)$, each link $e\in E$ is assigned a cost $c(e)\geq0$ and delay $d(e)\geq0$. A path $P$ from $s$ to $t$ has total cost $c(P) = \sum_{e\in P} c(e)$ and delay $d(P) = \sum_{e\in P} d(e)$. The DRCR problem is: $$\min_{P} c(P) \quad \text{subject to} \quad P \text{ is elementary } s\to t,\; L \leq d(P) \leq U$$ where $e\in E$0 is an end-to-end delay range. This problem is NP-complete; restriction to $e\in E$1 yields the classical constrained shortest path problem (Zhao et al., 2023).

DetNet 1+1 protection layers a further constraint: the selection of two Srlg-disjoint (shared-risk link group) $e\in E$2 paths $e\in E$3 with $e\in E$4 and $e\in E$5, essential to avoid egress packet reordering or buffer overflow during failover.

In decoupled architectures, the dropless guarantee depends on supporting on-the-fly routing with arbitrary dynamic policies, as captured in the forward progress predicate: If a policy-compliant path exists in the failure-pruned network $e\in E$6, all packets reach $e\in E$7 without mid-route discard due to topological change (Subramanian et al., 2019).

2. Algorithms and Mechanisms for Dropless Operation

2.1 Branch-and-Bound Routing: Pulse⁺ and CoSE-Pulse⁺

Pulse⁺ implements an aggressive branch-and-bound DFS for the DRCR problem. The algorithm prunes the search space using two main conditions:

• Feasibility pruning: If the cumulative delay of the partial path plus the minimum-delay suffix from the current node to $e\in E$8 exceeds $e\in E$9, the path is pruned.
• Optimality pruning: If the cumulative cost plus minimal remaining cost is no better than the current $c(e)\geq0$0, the path is discarded.

Pulse⁺ disables "dominance" pruning for $c(e)\geq0$1, preserving full optimal solution space coverage. Partial paths are extended in largest-delay-first order to encourage early discovery of low-cost feasible solutions, expediting global pruning. Pulse⁺ guarantees optimality upon termination (Zhao et al., 2023).

CoSE-Pulse⁺ generalizes this methodology for the Srlg-disjoint DRCR: it recursively partitions the Srlg inclusion/exclusion space based on detected conflict sets, using Active-Path Pulse⁺ and backup search to jointly construct feasible $c(e)\geq0$2 pairs. The algorithm applies conflict set-driven subproblem splitting, systematically excluding unsatisfiable combinations.

2.2 Data-Plane-Only Policy-Compliant Routing: D2R

D2R achieves dropless forwarding in a fundamentally different architectural setting: all routing computation occurs in the data plane. The architecture decomposes into three planes:

• Policy Plane: Operators define per-flow path policies. Policies and topology encodings are distributed to the data plane.
• Switch Control Plane: Minimal, only responsible for reacting to physical link state changes.
• Data Plane: On each packet, programmable ASICs execute a graph traversal (BFS for hop-optimality, or IDDFS for bounded state) starting from the current node, subject to the packet’s failed-link bit vector and switch-local port status. The first discovered policy-compliant path is written into the packet as a source route.

Failure propagation leverages Failure-Carrying-Packets (FCP): packets accumulate a bit vector of observed failures; as they traverse, each switch marks locally failed links, ensuring network-wide failure awareness without broadcast. The algorithm ensures that, as long as some compliant path exists, all packets are rerouted in hardware—dropping only those at in-flight failure instant (Subramanian et al., 2019).

2.3 Source-Controlled Alternate Paths: SlickPackets

SlickPackets eliminates forwarding plane dependency by embedding both the primary and alternate paths as a directed acyclic forwarding subgraph (FS) directly in the per-packet header. The FS provides, for each primary path edge, a disjoint alternate to the destination. The receiving router, upon detecting a failed next-hop, immediately transfers the packet to the embedded alternate, with no FIB recomputation or signaling required.

Header size is tightly bounded: for typical large ISP topologies (e.g., Sprint with 315 nodes), median header is $c(e)\geq0$3–$c(e)\geq0$4 bytes; even router-level topologies with $c(e)\geq0$5 nodes show $c(e)\geq0$6132 byte headers. For $c(e)\geq0$7-hop primary paths and $c(e)\geq0$8-length alternates, $c(e)\geq0$9 (Nguyen et al., 2012).

No cycles can arise due to the FS being a DAG, ensuring strict loop-freedom and that forwarding always terminates at $d(e)\geq0$0 within $d(e)\geq0$1 steps or is dropped only at a bona-fide disconnected failure.

3. Empirical Results and Performance Analysis

The empirical evaluation of dropless routing systems reveals substantial superiority in both failover time and completion rates:

Algorithm	DRCR Solves (10k nodes)	Srlg-Disjoint Solves (10k nodes)	Typical Runtime
Pulse⁺	100%	100% (CoSE-Pulse⁺)	<200 ms / <10 ms
Cost-KSP	~67%	<50%	Often >10 s, many time out
Delay-KSP	~15%	<50%	Often >10 s
Lagrangian-KSP	~95%	<50%	10–100× slower than Pulse⁺

Pulse⁺ and CoSE-Pulse⁺ maintain performance sufficient for real-time control-plane reaction at Internet scale (networks up to $d(e)\geq0$2 nodes), critical for large DetNet fabrics (Zhao et al., 2023).

D2R achieves sub-microsecond rerouting on physical failure, maintaining throughput at $d(e)\geq0$3 line-rate. Empirical loss is $d(e)\geq0$4 packet per failure, orders of magnitude superior to SDN or OSPF/BGP reconvergence mechanisms (200–1000 ms convergence, hundreds of packets lost). Recirculation requirements are modest (median $d(e)\geq0$5–$d(e)\geq0$6 for BFS, $d(e)\geq0$7–$d(e)\geq0$8 for IDDFS) even under multiple failures on 70-node topologies (Subramanian et al., 2019).

SlickPackets demonstrates that even for extremely large networks, the compactness of header encoding and immediate local reroute enable $d(e)\geq0$9 delivery under persistent single-link failures. Failure-recovery latency is governed by hardware link failure detection ($P$0 ms) rather than control-plane reaction (Nguyen et al., 2012).

4. System Design and Deployment Trade-offs

Dropless routing entails several context-dependent trade-offs:

• Computation Time: Branch-and-bound solvers grow super-polynomially in the worst case but are empirically practical for networks with tens of thousands of nodes. Algorithmic engineering (Pulse⁺ pruning, LDF order, hierarchical partitioning) is key (Zhao et al., 2023).
• Data-plane Overhead: D2R requires modest additional header and register state (two 64-bit registers per port; up to $P$1-bit per-packet source route; line-rate preserved for up to $P$2 recirculations) (Subramanian et al., 2019). SlickPackets’ encoding overhead is sub-30 bytes in practice but can inflate under $P$3-fail alternates (Nguyen et al., 2012).
• Policy Compliance: All approaches ensure that path policies (e.g., middle-box chain traversal, weighted next-hop) are strictly enforced, not merely best-effort, during failure scenarios.
• Legacy Interoperability and Incremental Deployment: Source-embedded alternate path systems (e.g., SlickPackets) may be difficult to incrementally deploy due to reliance on per-packet encoding and format support. D2R and Pulse⁺-style mechanisms typically require hardware or firmware updates in core routers or switches, but work with standard DetNet/Segment Routing abstractions.

A plausible implication is that dropless architectures impose requirements on network management tooling: policy distribution, topology encoding, and coordination between central controllers and data plane become critical for frictionless operation at scale.

5. Implications, Limitations, and Open Problems

Dropless routing realizes the theoretical ideal of convergence-free, lossless packet delivery through path failures, but several limitations and research directions remain:

• Multifailure Scalability: Encoding full $P$4-fail alternate sets can cause header bloat (Nguyen et al., 2012). Hierarchical partitioning and area-based routing mitigate recirculation and state growth (Subramanian et al., 2019).
• Control-plane Stress: Rapid failure bursts necessitate fast recomputation and distribution of alternate path (FS) structures or updated policies.
• ASIC/Hardware Feasibility: Fully hardware-based implementations are empirically demonstrated (D2R on Tofino), but alternative designs (especially source-code-based recoding in SlickPackets) may require further advances in network-processor platforms.
• Admission Control: To preserve dropless guarantees under heavy or changing load, admission control is necessary: e.g., link utilization costs are raised dynamically to prevent congestion violating per-path delay/jitter bounds (Zhao et al., 2023).
• Queue Management and Cycle Assignment: Dropless DetNet mandates careful queue-per-cycle allocation (e.g., CSQF), requiring compatible route computation and downstream cycle assignment (Zhao et al., 2023).

6. Research Landscape and Comparative Overview

Research on dropless routing spans centralized (optimal) path-computation engines (Pulse⁺/CoSE-Pulse⁺) (Zhao et al., 2023), data-plane-only policy-compliant reroute (D2R) (Subramanian et al., 2019), and source-controlled approaches (SlickPackets) (Nguyen et al., 2012), with complementary strengths:

Approach	Optimization Target	Failure Model	Policy Compliance	Hardware Dependencies
Pulse⁺/CoSE-Pulse⁺	Min-cost delay-bounded	Srlg, arbitrary	Centralized	Standard controller + ASICs
D2R	Policy-compliant path	Arbitrary	Per-packet	P4-programmable ASIC
SlickPackets	Source-specified FS	Single-link	Per-packet	Labeling, header parsing

The emergence of these techniques enables deployment of deterministic, lossless-by-design networks at Internet backbone scale and offers frameworks for future research in multi-failure protection, ultra-low-latency fabrics, and hybrid control/data-plane routing integration.