This paper presents RAPID, a resource allocation protocol for disruption-tolerant networks (DTNs) that optimizes specific routing metrics such as worst-case delivery delay or the fraction of packets delivered within a deadline. The key insight is to treat DTN routing as a resource allocation problem, translating the routing metric into per-packet utilities to determine how packets should be replicated in the system. RAPID evaluates these utilities at each transfer opportunity to decide whether to replicate a packet, aiming to maximize the utility of the network resources. The authors evaluate RAPID through a prototype deployed on a vehicular DTN testbed consisting of 40 buses and simulations based on real traces. The results show that RAPID significantly outperforms existing routing protocols in terms of several metrics, including average delay, worst-case delay, and the number of packets delivered within a deadline. For small loads, RAPID performs within 10% of optimal performance. The paper also discusses the design of RAPID, including its selection algorithm, inference algorithm, and control channel. The selection algorithm determines which packets to replicate based on their utilities, while the inference algorithm estimates the utility of packets given the routing metric. The control channel exchanges metadata about packet replicas and delivery status, improving routing decisions. Finally, the authors compare RAPID with other routing protocols, including MaxProp, Spray and Wait, Prophet, Random, and Optimal. They find that RAPID consistently outperforms these protocols, especially in terms of average delay and packet delivery rates. The paper also explores the impact of metadata exchange and the use of a global control channel on RAPID's performance.This paper presents RAPID, a resource allocation protocol for disruption-tolerant networks (DTNs) that optimizes specific routing metrics such as worst-case delivery delay or the fraction of packets delivered within a deadline. The key insight is to treat DTN routing as a resource allocation problem, translating the routing metric into per-packet utilities to determine how packets should be replicated in the system. RAPID evaluates these utilities at each transfer opportunity to decide whether to replicate a packet, aiming to maximize the utility of the network resources. The authors evaluate RAPID through a prototype deployed on a vehicular DTN testbed consisting of 40 buses and simulations based on real traces. The results show that RAPID significantly outperforms existing routing protocols in terms of several metrics, including average delay, worst-case delay, and the number of packets delivered within a deadline. For small loads, RAPID performs within 10% of optimal performance. The paper also discusses the design of RAPID, including its selection algorithm, inference algorithm, and control channel. The selection algorithm determines which packets to replicate based on their utilities, while the inference algorithm estimates the utility of packets given the routing metric. The control channel exchanges metadata about packet replicas and delivery status, improving routing decisions. Finally, the authors compare RAPID with other routing protocols, including MaxProp, Spray and Wait, Prophet, Random, and Optimal. They find that RAPID consistently outperforms these protocols, especially in terms of average delay and packet delivery rates. The paper also explores the impact of metadata exchange and the use of a global control channel on RAPID's performance.