Network Sciences: Coded Computing: A Transformative Framework for Tactical Wireless Edge Computing
Abstract
The use of compute-intensive machine learning and big-data analytics for processing a collection of raw data streams from distributed sources to enhance awareness of an environment and act on it is on the increase as smart things -- devices capable of a combination of sensing, communication, computation, storage, and actuation -- are becoming increasingly more prevalent, even on tactical battlefields. Unlike traditional cloud computing environments, however, tactical battlefield networks are characterized by significant challenges with respect to scalability due to (1) severe constraints on bandwidth and other resources including storage and energy; (2) high dynamics due to mobility and disruptions due to jamming and battle conditions. We propose to develop an innovative framework for large-scale tactical edge computing, named "Coded Computing", which brings new concepts from Shannon s information theory and coding to overcome these two scalability issues. Our solution is rooted in a strong theoretical foundation, is supported by real-world demonstrations, and can have a transformative technological impact. The proposed Coded Computing architecture essentially allows to reduce the bandwidth utilization of nodes at the expense of incurring some additional redundant computation, which is well suited to embedded battlefield devices. More fundamentally, as the network size increases, while the computation resources grow linearly with network size, the overall communication bandwidth is fixed and will become the bottleneck. Our proposed Coded Computing architecture allows to trade computation resources with communication resources in the network, hence removing the communication bottleneck and enabling a scalable design for tactical edge computing. Furthermore, the proposed Coded Computing architecture combines coding theory with distributed computing to inject computation redundancy in an unorthodox coded form (as opposed to the state-of-the-art replication approaches) to provide robustness to network dynamics and failures. To this end, we utilize more sophisticated coding techniques in distributed computing instead of the naive and inefficient repetition-coding scheme. This project, which consists of three research thrusts, takes a principled and foundational approach to providing a Coded Computing framework to tackle key bottlenecks in large-scale tactical wireless edge computing for machine learning applications. The first thrust focuses on the bandwidth bottleneck and develops Coded Computing approaches that are optimized based on inherent heterogeneity of network devices, the underlying network topology, and the structure of computations. We will also develop Coded Computing strategies to speed up iterative computations that arise in many machine learning problems, in particular iterative gradient methods that are currently the most widely used training algorithms. The second thrust focuses on the resiliency and robustness to network dynamics and failures and develops Coded Computing approaches that are scalable and dynamic for large classes of linear and non-linear computation tasks. We will also develop new distributed mechanisms for mitigating stragglers in iterative gradient-based machine learning algorithms. Finally, the third research thrust focuses on practical demonstration of the developed Coded Computing approaches via experiments over cloud networks and an edge computing testbed.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 14, 2019
- Source ID
- W911NF1810400
Entities
People
- Salman A. Avestimehr
Organizations
- Army Contracting Command
- United States Army
- University of Southern California