Migrating Large-Scale Air Traffic Modeling to the Cloud

Abstract

Coordinating nationwide air traffic flow is a large-scale problem. The modeling process generally involves analysis of massive flight data, and its optimization involves computationally expensive algorithms. This paper uses Hadoop MapReduce, a big data processing model, to facilitate air traffic flow modeling and optimization, where computationally intensive tasks are automatically spread to Hadoop clusters for concurrent executions. The overall wall-clock time of computation is reduced. A nationwide traffic flow management problem that has been previously studied was restructured under the MapReduce framework. The problem aims at minimizing flight delays while respecting system capacities. Due to its temporal and spatial scope, the size of this problem grows to an extent where it is too big to be solved on standalone computers. Lagrangian relaxation was applied to decompose the original problem into a collection of solvable subproblems. The optimization proceeds in two iterative stages: solving subproblems and Lagrange multiplier updates. These two processes are encapsulated in the mapper and reducer functions, respectively. As a result, the optimization is automatically scheduled to run in parallel tasks. The cloud-based air traffic modeling and optimization were validated through running nationwide air traffic optimization instances on a small Hadoop cluster with six nodes. The modeling processing is eight times faster and the optimization is 16 times faster than that running on standalone computers. Nomenclature A arr , A dep = set of links connecting origin or destination airport C s t = maximum number of aircraft allowed in sector s at time t C arr t, C dep t = airport arrival and departure capacity d k λ s t = objective of the kth subproblem f k t = departures into route k at time t K = set of routes in simulation and optimization n k = number of links on route k Q s i = set of links inside sector s i q k i t = outflow of link i at time t S = set of sectors in the National Airspace System s i = sector that link i lies in T = planning time horizon of air traffic optimization T k i = traversal time of link i on route k, min t = time step λ s t = Lagrange multiplier for sector s at time t Subscripts i = index of a link j = index of iteration s = index of a sector Superscript k = index of a route

Modeling, Optimization, and Operation of Large-Scale Air Traffic Flow Management on Spark

“…However, it requires extensive programming skills to implement multithreaded programming on a cluster, such as dealing with communication and synchronization issues. To overcome this limitation, a cloud computing framework with Apache Hadoop MapReduce was implemented to reduce the development workload from multithreaded programming [18], where Hadoop MapReduce is a software framework to process large-scale data in parallel on large cluster. With its built-in fault-tolerance capability, the MapReduce framework could be not only efficient but also robust.…”

“…A high-level description of the parameter setup process on Spark is provided next. We refer readers to [18] for detailed discussion of the parameters setup process with MapReduce. After the first stage, routes and links (segments of a route in a sector) information is stored in a database for the setup of subsequent LAM optimization.…”

“…This section presents NAS-wide instances of TFM problems that run on a Spark cluster. The goal is to compare the performance of the TFM problems on Spark with that on the existing Hadoop MapReduce framework [18]. To make the comparison meaningful, the simulation experiment was set up with the same hardware parameters.…”

A network based dynamic air traffic flow model for en route airspace system traffic flow optimization

“…[35][36]. The variables are reduced, but the dynamics of the air traffic flow inside sectors is missing.…”

A study of optimization of air traffic network and flight schedules

“…Sun [94] proposed a model using the Lagrangian method and parallel programming methods to reduce the computational time of the model. Programs can be run in the cloud so that it can eliminate hardware limitations.…”