Symbol-Synchronous Buses: Deterministic, Low-Latency Wireless Mesh Networking with LEDs

Abstract

Back to Top

1. Introduction

Attaining wire-like performance has long been a goal of wireless networking research. Indeed, successful technologies for the wireless collection of sensor data within the Internet of Things (IoT) are being marketed as achieving “wire-like reliability,”²¹ promising to replace costly and hard-to-maintain wired networks with a drop-in wireless solution. Under the hood, however, these networks operate quite unlike wires: the dominant type of network in these settings is the wireless mesh network (WMN), which builds on years of research examining how to reliably propagate data frames by fully receiving and then forwarding them from one node to another (“routing” by “store-and-forward packet switching”), and how to prevent different nodes from transmitting concurrently on the same channel in order to avoid destructive interference (“medium access control” (MAC)). Current IoT protocol stacks hence instantiate routing protocols, which establish the paths along which messages should be forwarded, as well as MAC protocols, which enforce schedules that determine when a node’s radio should transmit a data frame, that is, a sequence of symbols that communicates data to the next hop in the path.

Store-and-forward routing and transmission scheduling result in end-to-end latency (i.e., the time it takes for a frame to be transmitted from a sending to a receiving interface across the network) that is far larger and much more variable than for wired industrial control networks.^{2, 13} Since low latency and jitter (i.e., variance in latency) are paramount for the latter type of network, the applicability of wireless meshes in latency-sensitive or event-driven scenarios, such as real-time control or robotics, remains limited.¹⁶ The use of low-power wide area networks (LPWANs) instead of WMNs does not offer a solution either: LPWANs may eliminate multi-hop routing, yet they do so at the expense of throughput and reliability.¹

If wireless networks are to one day support the variety of embedded applications that still depend on wires, innovative solutions will be needed. This research highlight describes one such solution, the essence of which is to eliminate store-and-forward packet switching and to replace it with a novel networking paradigm, the symbol-synchronous bus, which makes a wireless mesh network behave like a single wire. In such buses, nodes do not wait for the complete reception of a data frame, or even a single symbol within it, before re-broadcasting it to their neighbors. In pursuit of this vision, we introduce the ZERO-WIRE platform, a hardware and software prototype of a symbol-synchronous bus that uses optical wireless communication (OWC), that is, free-space signaling between commonly available light sources and photodetectors. OWC is a natural medium on which to implement the prototype due to the abundance of largely unregulated bandwidth in the optical part of the electromagnetic spectrum, its remarkable interference characteristics, and the simplicity of OWC transceivers,¹⁸ which facilitates low-level research.

The remainder of this paper explains in more detail how symbol-synchronous buses and ZERO-WIRE achieve a performance envelope that more closely resembles wired bus networks than wireless meshes. Section 2 provides the necessary background on store-and-forward IoT networking, wired buses, and OWC. Next, Section 3 discusses how some of the introduced concepts combine to explain the operation of a symbol-synchronous bus. Section 4 introduces ZERO-WIRE’S design, and Section 5 evaluates its performance under laboratory conditions. Section 6 concludes the paper and elaborates on future research opportunities.

Back to Top

2. Background

This section provides the background information required to understand the design of ZERO-WIRE. Section 2.1 details the limitations of store-and-forward packet switching for the IoT. Next, Section 2.2 discusses attractive features found in wired buses. Section 2.3 then reviews the fundamental differences between optical and radio communication.

 2.1. Store-and-forward packet switching

Current IoT networking technologies struggle to support low-latency applications, particularly in dense deployments. Contemporary mesh networks introduce 100–1000s of milliseconds of latency and 10s of milliseconds of jitter in typical configurations,¹³ yet tactile feedback or robotics may require deterministic end-to-end delays of under 1 ms.²⁰ Similarly, end-to-end and single-link throughput often differ by more than one order of magnitude, unless a specific multi-hop path is optimized to suit well-known traffic demands.^10,12 While the performance characteristics of specific traffic flows can hence be improved by, for example, favoring them over others in transmission schedules,⁶ such monopolization of network-wide resources is hard to reconcile with applications involving bursty, dynamic, in-network, any-to-(m)any, or event-driven traffic.⁹

Striving for IoT networks that better suit latency-sensitive applications, researchers have examined Synchronous Transmission (ST).²³ ST systems do not prevent nodes from interfering, but correctly decode frames in spite of concurrent channel access by synchronizing transmitters. Glossy,¹¹ a seminal example, lets an initiator node broadcast a frame through a mesh network. Nodes that receive the initiator’s frame repeat it while synchronizing their transmissions down to 0.5 μs. Such synchronization enables error-free demodulation of transmitted frames with high probability due to modulation-specific mechanisms that are commonly misinterpreted as constructive interference.¹⁵ The process then repeats until it has flooded the entire mesh network.

ST enables end-to-end communication at a few milliseconds of latency¹¹ and a higher sustained throughput than conventional approaches.²³ However, it remains a store-and-forward solution: Nodes receive and decode all symbols that make up a data frame before transmitting themselves. Latency is therefore heavily impacted by network diameter, frame length, and the repetition of preambles and headers.¹¹ When multiple initiators are present or multiple packets need to be sent, distinct traffic flows should also be prevented from interfering, which reintroduces the need for medium access control.²³

 2.2. Bus networks

Wired buses are conceptually simpler than wireless mesh networks: the former connect all nodes through a single, shared set of wires. These wires provide fast and deterministic half-duplex links that are isolated from their environment and offer all nodes a consistent view of the network: any signal written to the bus effectively reaches all nodes instantaneously (i.e., after a propagation delay that is small relative to the symbol duration). As all nodes on the bus see all traffic, they can refrain from interfering with ongoing transmissions and may sense the channel to detect collisions while transmitting. In some buses, the latter mechanism enables priority-based arbitration,^2,14 which exploits the dominance of some symbols over others to non-destructively resolve contention: if multiple nodes transmit concurrently, nodes suppress low-priority traffic before it corrupts a high-priority frame, which is received without delay or coordination between nodes, as if the other colliding frame had never been transmitted. This results in perfect network utilization: the network’s performance is independent of the number of nodes, their density, or the required traffic patterns. The simplicity, performance, and determinism of wired buses have made them a popular communication interface for both resource-constrained electronic peripherals (SPI, I²C, 1-Wire) and real-time industrial control (Profibus, CAN).^2,14

 2.3. Optical wireless communication

LEDs and photodetectors communicate data by modulating the intensity of light at the transmitter and converting the observed intensity into an electric current at the receiver. Such intensity modulation with direct detection (IM/DD)¹⁸ differs from conventional radio modulation, since it embeds information in the power envelope of a signal, rather than in the low-level properties of the underlying electromagnetic waves, such as their phase, amplitude, or frequency. This difference affects the way in which concurrent signals interfere. Intuitively, if two LEDs are modulated to transmit the same signal at the same time, the intensity of the light as seen by a detector is roughly twice that of one LED. The same setup with two radios does not double the amplitude of the transmitted electromagnetic wave, since carrier frequency offsets between transmitters induce beatings in the resulting signal¹⁵ and signal-to-noise ratios vary wildly with the receiving radio’s position due to the location dependency of the signals’ phase offset.²² LED-based signals are far less susceptible to these phenomena because LEDs are incoherent emitters: the coherence time of the waves that make up LED light, that is, the period of time during which the phase of a wave remains consistent, is orders of magnitude smaller than the sampling interval of any reasonable intensity detector.⁸ This means that interference between IM/DD signals of incoherent emitters can be understood, mitigated, and designed for at the (de)modulation level:¹⁹ the resulting signal is the sum of the power envelopes of the interfering signals, and it is not meaningfully affected by out-of-reach low-level phenomena.

Back to Top

3. Paradigm

 3.1. Symbol-synchronous relaying

Symbol-synchronous buses are wireless mesh networks that, like wired buses, offer sub-millisecond end-to-end latency for small frames, microseconds of jitter, and perfect priority-based network utilization. As illustrated in Figure 1, attaining such features in a mesh network requires parallelizing retransmissions between source and destination nodes as much as possible, while avoiding MAC overhead between possible forwarders. More precisely, relay nodes relay any symbol immediately after observing it, without coordinating with other nodes, and without making forwarding decisions based on the contents of the message. Peripherals, such as sensors and actuators, are not required to relay themselves and may instead connect to a mesh of full-duplex relays, referred to as the network’s canopy.

Figure 1. Symbol-synchronous systems mimic the time behavior and global consistency of wired buses, while store-and-forward systems do not.

The key innovation of symbol-synchronous buses lies in the way relay nodes avoid feedback loops between their concurrent retransmissions. As illustrated in Figure 2, nodes relay incoming symbols with a relay offset time r such that the time offset q between a signal’s first arrival and any relayed version perceptible to the same node is smaller than the symbol duration s. As shown in the figure, this constraint allows nodes to distinguish symbols that they still have to relay from symbols they have already forwarded and should not be relayed again by the same node. The point in time s after a symbol’s first arrival thus becomes a demarcation point between old and new data, so nodes do not need message identifiers or addressing information to make forwarding decisions and do not require multiple transmission bands. Equivalently, a sending node maintains a symbol duration that is slightly longer than necessary for a receiver to decode the symbol: in the figure, the peripheral’s value for s leaves time for relay node “A” to relay the incoming symbol. This time constraint guarantees at least one moment during the transmission of every symbol where nodes in a one-hop neighborhood are symbol-synchronous: they all transmit and receive the same symbol at the same time.

Figure 2. Symbol-synchronicity prevents nodes from re-relaying signals they already relayed.

 3.2. Multi-hop propagation

As shown in Figure 1, the result of symbol-synchronicity is virtually deterministic timing behavior similar to that of wired buses. For both types of network, latency is determined only by the transmission time of a single frame and a relatively small propagation delay. Specifically, for a symbol-synchronous bus, end-to-end latency l for an N-symbol frame is bounded by its transmission time N · s, the per-hop relay offset r, and the network diameter d: Ns ≤ l ≤ Ns + rd, where r < s and typically rd << Ns, as will be demonstrated later in this paper. As with store-and-forward solutions, latency increases linearly with the number of hops separating the initiator and receiver, yet every hop adds only r < s microseconds instead of the transmission time of a full frame. This approach subsumes routing and medium access control among forwarders: the entire canopy transmits concurrently, so the bus simultaneously connects source and destination(s) through all existing paths.

As illustrated in Figure 3, individual symbols move away from the node that initiated the transmission in a wavelike fashion. Nodes propagate the wave by switching from the transmission of one symbol to the next. The symbol-synchronicity time constraint essentially guarantees that no node observes two wave fronts at the same time, thus disambiguating which symbol needs to be transmitted. Symbol-synchronicity hence is a local concept: it considers the signals transmitted and received by a node and its immediate neighbors. An initiator is free to start transmitting the next symbol as long as it can guarantee that doing so will not cause the new wave front to collide with old ones. This implies that, in some cases, a stricter time constraint is needed. Concretely, since the transmission of subsequent frames may be initiated by different nodes, the end of a frame should in general be indicated in a globally symbol-synchronous way. In other words, to prevent wave fronts belonging to different frames from colliding, all nodes must agree that a frame’s transmission is over before the next one can be sent. Analogously, if we hope to implement a symbol-synchronous bus that provides non-destructive arbitration by having a symbol transmitted by one node dominate those sent by contending transmitters, all nodes must agree on which symbol is being transmitted before the next one is sent (i.e., rd < s). In both cases, propagation delays impose a maximum symbol rate B = 1/s at which the network can operate reliably.

Figure 3. Symbols ripple through the network like a wave.

 3.3. Inter-signal interference

To successfully propagate and decode information, nodes must discern a single initiator’s signal among a collectively much stronger set of signals that are emitted by neighboring relay nodes and that slightly lag behind it. More precisely, as shown in Figures 1 and 2, nodes must recognize the initiator’s next symbol while it is being interfered with by many more transmissions of the previous symbol. Since the network in no way constrains what forwarding paths will be used, nodes do not know how many concurrent transmissions there are, nor their exact time offsets or signal strengths. OWC, or intensity modulation of an incoherent transmitter, offers a solution: since interference patterns can be predicted reliably from the choice of symbols and their time offsets alone (cf. Section 2.3), a modulation scheme can be designed to help nodes infer what the initiator’s signal looks like despite their imperfect understanding of many more interfering signals. Nodes in a symbol-synchronous bus therefore rely on a transceiver that is specifically designed to accomplish this task, rather than one that tries to cancel out interference for an already established modulation scheme, as is done by several state-of-the-art techniques⁷ (we refer the reader to the full version of this paper for an in-depth comparison). The primary goal of symbol-synchronous modulation is hence rapid forwarding in mesh topologies, rather than efficient communication of information across a single hop. Symbol-synchronous relaying aids this approach by constraining what symbols can be transmitted concurrently and how far apart they can be in time.

Back to Top

4. Design and Implementation

 4.1. Transceiver logic

The ZERO-WIRE platform addresses the challenges identified in Sections 3.1–3.3 for a simple modulation scheme based on the On-Off Keying (OOK) of LEDs, which has an extremely low barrier to entry for resource-constrained devices. All that is required for a node to transmit data through the network is to blink an LED within reception range of the canopy at an arbitrary but fixed modulation speed. Still, many other IM/DD modulation schemes have been proposed,¹⁸ and exploring their trade-offs in symbol-synchronous networks is a target of our future work.

Nodes participate in a ZERO-WIRE network by transmitting and receiving data using a transceiver that implements symbol-synchronous modulation logic. A schematic overview of that transceiver is provided in Figure 4. The transceiver implements a real-time control loop that repeatedly updates a set of interconnected state machines in sequence while interfacing with a photodetector, LED(s), and an external processor. The LEDs enable the modulation of data, the photodetector enables the sampling of ambient light intensity, and the state machines process the samples to recover data. The transceiver reads frames to be sent from—and writes frames that were successfully received to—a set of buffers that can also be accessed by the external processor. The control loop interval i is fixed: every i microseconds, a node samples the ambient light level exactly once and performs all processing steps associated with the new sample. If i is sufficiently small, this approach helps to guarantee that nodes recognize and start relaying a new symbol at a small and predictable time offset r.

Figure 4. Overview of a ZERO-WIRE transceiver.

The transceiver’s design parametrizes s in terms of the number of samples per symbol and minimizes a fixed r. For a sufficiently large s, the symbol-synchronicity time constraint is then satisfied by design. As will be demonstrated in Section 5, the maximum symbol rate at which the network remains operational may therefore be obtained empirically for any given topology. In essence, an iteration of the control loop comprises the following six steps. We refer the reader to the full version of this paper for an in-depth discussion of the mechanisms that implement each step.

1. The transceiver samples a photodetector, retrieving a number that represents the current intensity of ambient light.
2. A first state machine reasons on a recent history of samples to detect edges, that is, one or more neighboring nodes that change the state of their LED(s) in response to the initiator’s signal.
3. A second state machine decides whether the transceiver’s own LED(s) should be on or off during the next i microseconds. That decision considers the output of the first state machine and the presence of data to be sent in the downstream buffer. If arbitration requires a node to suppress its ongoing transmission, this state machine informs the external processor of that event.
4. The control loop drives the transceiver LED(s) on or off, based on the decision made in the previous step.
5. A third state machine keeps track of the time (i.e., the number of loop iterations) between recent edges and, leveraging ZERO-WIRE’S frame structure, recovers transmitted data.
6. The control loop decides whether the current loop iteration marks the end of a (valid) frame to be passed to the external processor. That decision also informs parameter setting for ZERO-WIRE’S edge detection logic.

 4.2. Implementation

ZERO-WIRE builds on the DenseVLC⁴ hardware platform, which provides a physical-layer transmitter and receiver front-end for OWC. We orient the LED and photodiode on the transceiver in the same direction, with the LED outside the field of view of the photodiode, to prevent transmitted signals from saturating receiver amplifiers. This positioning enables the deployment of relay nodes such that they receive signals through the latter’s reflection off walls and floors.

We implement the ZERO-WIRE transceiver in software using ~500 lines of C code run on an STM32G474 microcontroller, which interfaces with the DenseVLC front-ends. We have also ported this implementation to the BeagleBone Black board.³ Both versions are interoperable and attain the same performance levels. The clock speed of the processor core that implements the transceiver imposes a minimum control loop interval i = 2.5 ms.

Back to Top

5. Evaluation

 5.1. Test bed setup

This section evaluates ZERO-WIRE in terms of end-to-end network performance. Since large-scale OWC test beds that could support our evaluation remain unavailable,⁵ we created our own test bed of 25 ZERO-WIRE relay nodes in a laboratory environment. As shown in Figure 5, the nodes are all positioned in the same plane on a 5-by-5 grid and oriented such that their LEDs point toward a grey-painted surface, connecting them through their signals’ diffuse reflection. Such a planar topology, which mimics a deployment of ceiling-mounted light bulbs that communicate via their signals’ reflections off walls and floors, is both practical and common in contemporary OWC research.⁴ The dimensions of the test bed, also shown in the figure, and the relative positions of nodes within it, were chosen to enable an examination of how the network’s performance characteristics scale with parameters of the network topology in a controlled setting (cf. Section 5.3). In particular, the setup ensures the number of columns in the grid of nodes corresponds to the network diameter d. Nodes in different columns are positioned one hop apart, so a signal from a node in one column reaches those in neighboring columns, but not those farther away. By contrast, the number of rows in the grid corresponds to the network’s density k, since nodes in the same column form a clique, that is, they correctly receive each other’s signals without requiring intermediate relays. Higher values for k thus provide more links between nodes in neighboring columns. We position the test bed such that, on one side, it is exposed to a variety of natural and artificial light sources present in our offices over the business days during which the experiments were performed. The three other sides are opaque and diffusely reflect optical signals.

Figure 5. Test bed setup.

 5.2. Performance characteristics

Scenario. We study ZERO-WIRE’S performance characteristics by replicating traffic patterns resembling those used to interact with SPI- or I²C-based peripherals. Specifically, we instruct pairs of nodes to exchange frames in a request-reply pattern. One node sends a frame carrying L bits of data through a ZERO-WIRE bus configured to run at speed b; the receiving node replies with another frame of length L immediately after it detects the end of the request frame. This process then repeats, enabling a study of end-to-end latency, reliability, and goodput. For the purposes of this section, latency is the delay between the time at which the requesting node completes its request by observing its own message on the bus and the time at which it completes reception of the corresponding reply. Reliability then is the proportion of reply frames correctly received by the request node within all reply frames that were sent. Goodput is the number of bits in correctly received frames divided by the time span of the repeated request–reply process from the request node’s perspective. We arbitrarily limit the number of consecutive requests to 10 and subsequently restart the process with a new random pair of nodes, thus eventually collecting information on the performance of every end-to-end path within the network. This variation in traffic flows does not involve network-wide reconfiguration: no node is aware when another node will transmit; any node can decide to use the full capacity of the bus at any time without prior or consistent resource allocation. Metrics reported in this evaluation are averaged over all such end-to-end traffic flows.

Latency. Figure 6 shows end-to-end latency for ZERO-WIRE as a function of frame length. Latency is shown to be proportional to the speed of the bus and the amount of data to be transmitted. As shown in the inset, ZERO-WIRE enables sub-millisecond latency as long as the amount of data to be transmitted is sufficiently small relative to the speed of the bus, which in this topology enables the communication of a 2-byte payload across a four-hop network in one millisecond. We discuss the implications of using such small frame sizes in Section 5.4. As shown in this figure and explained in Section 4.1, only certain discrete settings for the bus speed b are possible, since symbol duration can only be configured in terms of the number of samples per symbol.

Figure 6. End-to-end latency.

In a ZERO-WIRE network, the key lower bound on latency is the fixed time cost of transmitting per-frame headers, as evidenced by the nonzero y-axis intercept of the lines in the figure. The rest of the latency profile is driven by payload size only. The obtained latency numbers are virtually deterministic: for any given bus speed and frame length, they vary by no more than tens of microseconds across traffic endpoints or experimental replicates. Variance in latency due to hop-by-hop relaying across paths of different lengths is negligible, since the main sources of such variance are propagation delay between neighboring nodes (r ≈ i = 2.5 μs per hop) and differences in the positions of nodes within their control loops.

Reliability. Figure 7 illustrates ZERO-WIRE’S reliability profile. For messages no longer than a few bytes and up to a bus speed of 20kbps, the end-to-end frame delivery ratio is practically constant at 99%, after which reliability collapses as the symbol-synchronicity time constraint cannot be consistently met (cf. Section 3.1). Reliability primarily depends on frame length and favors short frames: delivery ratios decrease by roughly one percentage point when using 128-bit instead of 32-bit frames. Contention, on the other hand, has no significant effect. When configuring two nodes to reply to a request, the higher-priority frame is received as reliably as frames that are not contended with.

Figure 7. End-to-end reliability.

We observe that corrupted frames are typically received as fragments: nodes must decide how to interpret a symbol while that symbol is being transmitted and cannot wait for the completion of a frame to mask transmission errors before relaying information. While future research may improve ZERO-WIRE’S reliability with more advanced signal processing, this observation suggests that the relationship between frame length and reliability is inherently stronger for a symbol-synchronous bus than for conventional networks, and thus motivates shorter frame lengths.

Goodput. Figure 8 details end-to-end goodput for a ZERO-WIRE network. Goodput increases with the data rate up to b = 20kbps, that is, up to the point at which the symbol-synchronicity constraint breaks and the network becomes unreliable. Goodput is subject to a trade-off: longer frames amortize per-frame overhead over a larger amount of payload data, but are also considerably more likely to be corrupted. Goodput-optimal frame lengths are situated between 128 and 256 bits and result in a goodput of 19kbps at a raw data rate of 20kbps. In that setting, ZERO-WIRE utilizes 95% of its theoretical throughput capacity for application data. Shorter frames enable low-latency applications, but incur a considerable goodput penalty: When using 8-bit frames, the network attains only 63% of its throughput capacity. Larger-than-optimal frames cause slow goodput loss, so they may be better sent as fragments.

Figure 8. End-to-end goodput.

 5.3. Scaling behavior

Scenario. We evaluate scaling behavior by removing rows and columns of nodes from the deployment in Figure 5. Removing a column reduces the number of nodes in the network by five and decreases its diameter d. Due to the parameters of our test bed, such diameter scaling can continue until only a five-clique of nodes remains. Density scaling, that is, removing a row of five nodes, leaves the network’s diameter unchanged, but decreases its density until only a five-node, four-hop network remains. For all ten topologies obtained this way, we repeat the analysis from the previous section.

Results. For a given bus speed, end-to-end latency is practically unaffected by the network’s diameter or density. As discussed in Section 5.2, such latency is virtually deterministic by design: variance in latency is on the same order as the granularity with which our test bed can measure latency (i.e., 10 s of μs), rendering further analysis moot.

By contrast, reliability, and therefore goodput, is substantially affected by parameters of the network topology. Figure 9 provides an overview of goodput for each of the network topologies under consideration. For the purposes of this figure, we report optimal goodput, as described in Figure 8. When scaling the test bed’s diameter, goodput tends to decrease as the number of nodes increases, falling from 25.4kbps in a five-clique to the 19kbps observed in the previous section for a 25-node network. When scaling network density, however, good-put increases from 16.5kbps to 19kbps when going from a minimally to a maximally dense configuration. In neither scaling scenario does the presence of contention have a considerable impact.

Figure 9. Effect of network topology size on goodput.

Scaling network diameter. Figure 10 reveals why scaling the network diameter decreases goodput by plotting frame delivery ratios in function of bus speed, for varying diameter d and 128-bit frames. When increasing the number of hops in a network, the reliability curve established in the previous section shifts to the left. The highest data rate setting at which the network operates reliably thus decreases. The goodput-optimal data rate setting hence goes from 28.6kbps to 20kbps as d goes from 1 to 5, explaining most of the goodput loss. The inset demonstrates that adding nodes to the network also has a measurable effect on reliability that is independent of this left shifting, introducing roughly a percentage point of additional frame loss in our scenario.

Figure 10. Effect of network diameter on reliability.

The root cause for the shifting of the reliability curve is described in Section 3.1. When adding hops to the network, a ZERO-WIRE node becomes exposed to sporadic interference by nodes whose signals are further offset in time. To maintain symbol-synchronicity, nodes therefore need to rely on a longer symbol duration and hence a lower data rate. In ideal circumstances, the time offset between signals emitted by neighboring nodes corresponds to one control loop interval i = 2.5 μs, making the curve shift to the left by one discrete speed setting (i.e., the minimal value for s/i for a reliable network increases by one). As shown in the figure, this means throughput capacity (i.e., optimal b, the point of network collapse in terms of bus speed) decreases sublinearly with increasing d, since b is inversely proportional to the number of samples per symbol (i.e., s/i). In a contention-free scenario, the need to decrease bus speed due to diameter scaling should eventually halt, since symbol-synchronicity is a local constraint (cf. Section 3.2). We do not yet observe this phenomenon in our four-hop test bed.

Scaling network density. Figure 11 illustrates why increasing network density improves goodput. For varying density k, the figure plots frame length against network utilization, that is, the ratio between observed goodput and the optimal raw data rate b. For low density, ZERO-WIRE’s reliability decreases more rapidly with increasing frame length. The net result is that, for k = 1 and k = 2, the tails of the goodput curves are tilted downward relative to more dense settings. Maximum goodput is then achieved for frame lengths of 32 and 48 bits, far smaller than the 128–256 bits observed for sufficiently dense topologies. Network utilization therefore tops out at around 80% instead of 95%. The optimal bus speed setting, used to obtain these results, varies between b = 20kbps and b = 22.2kbps without a clear trend. The apparent decrease in goodput shown in Figure 9 for density scaling from 20 to 25 nodes hence does not allow for generalized conclusions on network performance beyond a 25-node scenario: It is the result of a variation in bus speed (b = 22.2 to b = 20kbps), which cannot be masked by the small increase in network utilization from k = 4 to k = 5.

Figure 11. Effect of network density on network utilization.

Mutiple phenomena may explain the effect of network density on the relation between frame length and reliability. For example, synchronous intensity modulation by multiple nodes results in larger intensity variations than would be caused by any single node (i.e., signals “interfere constructively”). Additionally, networks with higher density exhibit a larger number of redundant paths: if a node fails to detect an edge in a signal, one of its neighboring nodes may still detect and subsequently relay it to the former node, masking the initial failure.

 5.4. Discussion

Performance envelope. ZERO-WIRE unlocks new performance points for wireless mesh networks: its end-to-end latency is virtually deterministic and has a lower bound that is smaller than conventional wireless meshes. Yet, its 99% reliability is considerably lower than the five nines claimed for state-of-the-art scheduled sensor networks.²¹ From a goodput perspective, ZERO-WIRE’s performance is roughly on par with radio-based low-power wireless mesh networks that are not optimized for a specific traffic pattern,¹⁰ while the platform, at the cost of inhibiting co-existent traffic flows, provides a mechanism for contention resolution with no latency cost. To realize these novel possibilities, ZERO-WIRE frames may have to be considerably smaller than the already small frames found in the current IoT networks.¹²

The above comparisons do not present a complete image, however. Mature networking technologies actively trade off certain performance metrics against others through, for example, retransmission or fragmentation. A future protocol on top of ZERO-WIRE’s low-level architecture must explore these trade-offs. Still, we believe our results point toward a platform that enables wireless interaction with remote devices that is more reminiscent of I²C-like buses than of conventional wireless networks.

Range. ZERO-WIRE, like other OWC platforms,¹⁸ is a short-range technology: initial measurements indicate that the reliability of the current prototype breaks down rapidly beyond 2 m of (line-of-sight) range for a single pair of nodes, and that low bus speeds are more robust over longer distances. Quantifying range as a single number is problematic because of its relationship with transmission power, available energy, topology density (cf. Section 5.3), and directionality of the optical front-end. Moreover, the impact of a network’s diameter—and thus that of a node’s range—on end-to-end performance may be considerably smaller than for conventional architectures. A complete investigation of ZERO-WIRE’s transmission range therefore requires the design of low-power peripheral hardware and deployment scenarios fine-tuned to the type of network it introduces.

Transceiver implementation. The phenomena that drive ZERO-WIRE’s performance are contingent on the number of intensity samples per symbol and the time it takes to process such a sample. With a control loop requiring roughly 500 single-cycle instructions at 200MHz, that time currently cannot be reduced below 2.5 μs. Moreover, stalls of the processor implementing the transceiver occasionally cause control loop iterations to finish late and may consequently lower reliability. In our vision, future versions of the ZERO-WIRE transceiver should therefore be implemented as a dedicated hardware module instead of a software-defined platform.

Back to Top

6. Conclusion and Outlook

This research highlight described the symbol-synchronous bus, a networking paradigm that provides end-to-end connectivity across a wireless mesh network and breaks conventional abstractions of store-and-forward packet switching, medium access control, and routing. The ZERO-WIRE platform implements this paradigm using optical transmissions, and empirical results in a laboratory setting demonstrate how a ZERO-WIRE mesh attains performance characteristics and interaction patterns that, until now, have been impossible to deliver using wireless mesh networks. In particular, a first ZERO-WIRE prototype has been shown to provide deterministic sub-millisecond end-to-end latency for two-byte frames, priority-based contention resolution without collisions, 19kbps of goodput, and 99% reliability. Nodes can access the network’s resources at a very low barrier to entry: transmitting data through the network is as simple as blinking an LED at regular intervals.

Several research tracks identified in this paper may enable the improvement of our initial evaluation results. As ZERO-WIRE currently relies on an unoptimized hardware platform, our future work will naturally investigate performance improvements that might be obtained with tailor-made hardware for both relay nodes and peripherals, while also examining the potential of symbol-synchronous transmission in other regions of the electromagnetic spectrum. Additionally, we aim to characterize these platforms in terms of evaluation criteria that are not (fully) addressed in this paper, such as transmission range and power consumption. Future protocol design efforts should also reconcile ZERO-WIRE with duty-cycled low-power operation of peripherals, for example, through light-based active wakeup,¹⁷ as well as explore how to best exploit the performance envelope unlocked by the platform to enable a novel class of wireless embedded applications.

Back to Top

Acknowledgments

This research is partially funded by the Research Fund KU Leuven, Jonathan Oostvogels’ Ph.D. fellowship (11H7921N) of the Research Foundation–Flanders (FWO), and the FWO D3-CPS project. The authors thank Jona Beysens and Qing Wang for their support regarding the DenseVLC hardware platform.