Joint Task Offloading and Routing in Wireless Multi-hop Networks Using Biased Backpressure Algorithm

TL;DR

This work tackles joint offloading and routing in wireless multi-hop networks where interference and time-varying queue states complicate decisions. It introduces an extended graph with a virtual sink for each task type and per-commodity queues, enabling computation to be treated as virtual-link traffic and enabling the use of Shortest Path–biased Backpressure (SP-BP) routing in a distributed fashion. A BP-compatible job scheduler and a four-step SP-BP routing/scheduling procedure yield a throughput-optimal, fully distributed solution that balances load, routes effectively, and minimizes makespan. Numerical results on 100-node networks show that the proposed joint_SP-BP scheme outperforms separated BP offloading and LP-based approaches, while mitigating the last-packet problem and scaling well with network size and traffic load.

Abstract

A significant challenge for computation offloading in wireless multi-hop networks is the complex interaction among traffic flows in the presence of interference. Existing approaches often ignore these key effects and/or rely on outdated queueing and channel state information. To fill these gaps, we reformulate joint offloading and routing as a routing problem on an extended graph with physical and virtual links. We adopt the state-of-the-art shortest path-biased Backpressure routing algorithm, which allows the destination and the route of a job to be dynamically adjusted at every time step based on network-wide long-term information and real-time states of local neighborhoods. In large networks, our approach achieves smaller makespan than existing approaches, such as separated Backpressure offloading and joint offloading and routing based on linear programming.

Figures (3)

• Figure 1: Graph modeling: (a) Task offloading in a wireless multi-hop network where a client may have two types of tasks. (b) Extended graph and per-commodity queues for the example in (a). (c) Job scheduling for two types of jobs with equal priority at a computing node $i$ with four processors, the computing capacity of a server is modeled as two virtual links, with link rate equal to the corresponding service rate.
• Figure 2: Numerical results on networks of 100 nodes: (a) Median makespan under different schemes as a function of traffic load, and (b) Median number of hops a job traveled before being processed, under the setting of two types of tasks. (c) Median makespan under different formulations as a function of traffic load under a single type of task. The bands indicate 25 and 75 percentiles, lines for median.
• Figure 3: An exemplary network with $100$ nodes, where node color indicates its role and node size represents computing capacity.