Understanding Concurrent Transmission: a Generic Platform for Complex IoT Applications

Current resource-constrained, low-rate wireless networks face prohibitive latency: they lack support for the sub-millisecond, multi-hop interactions between densely deployed distributed sensors and actuators that would enable or greatly benefit envisioned cyber-physical applications, such as real-time feedback control and privacy-preserving distributed inference. This dissertation argues that that gap might be overcome by re-thinking how wireless transceivers use symbols, the smallest units of information in digital transmissions. In particular, this dissertation introduces and prototypes the symbol-synchronous wireless network architecture, in which the use of wireless symbols is reconsidered in three ways.

First, this dissertation introduces symbol-synchronous modulation, a physical-layer technique wherein nodes strive to relay symbols within a small fraction of the symbol duration after observing them, rather than accurately recovering sequences of symbols from a given origin. By exploiting incoherent physical signals as well as a novel notion of synchrony, this approach causes an end-to-end communication channel to emerge that enables nodes across a mesh network to communicate with the contention-agnostic end-to-end latency that was formerly reserved for wired fieldbuses. Experiments on a 25-node wireless optical test bed in a laboratory setting demonstrate, for example, the propagation of a two-byte frame across four hops in less than one millisecond.

Secondly, this dissertation introduces collages, a link-layer technique with which symbols interact over the air to compute a series of maxima over small integers that are concurrently transmitted by several nodes. By adjusting the distribution of integers with respect to the number of concurrent transmitters, collages can be made to convey aggregate properties of distributed data, rather than that data itself. Bayesian analyses show that the collection of such aggregates, when interpreted relative to prior application knowledge, accelerates real-time classification: the same level of classifier performance can be attained in 15–50% less time, when compared to conventional network stacks.

Thirdly, this dissertation introduces ambiguous transmissions, an application- layer technique that adjusts the mapping of sensor data to and from the symbols comprised in collages. Adjusting said mapping enables the estimation of a collection of histograms that, lossily and in context-dependent ways, describes the network-wide data distribution, rather than enabling the reliable and context-free recovery of node data. Said context is provided by, for instance, developer-provided hints about application semantics, such as the available latency budget and the amount of measurement noise. Experiments reveal that such transmissions permit the tuning of networks for latency-bound rather than reliability-bound performance, accelerating data collection for high-noise wireless sensing problems, such as those involving locally differentially private sensor data, by 25–75%.

The result of this threefold approach is a network architecture wherein its components interact with one another through low-latency rather than reliable interfaces, repeatedly deferring failure handling to components that operate at higher levels abstraction. As such latency-first network design reveals considerable follow-up research opportunities, this dissertation concludes by sketching future directions for the development of mature symbol-synchronous hardware, as well as for the more general problem of devising engineering approaches that enable the construction of resource-constrained wireless communication systems in ways that are demonstrably better aligned with application semantics.