1. INTRODUCTION

Optical wireless networks have the potential to serve space–space, space–terrestrial/aircraft, and aircraft–aircraft data centers and metropolitan area networks. This paper addresses architecture features necessary for these networks. Free-space optical (FSO) networks have two dimensions that are not encountered in fiber networks and those are (1) the ability to connect without predeployment of infrastructures and (2) the ability to reconfigure its connection topology by beam steering in time scales of milliseconds to seconds to adapt to traffic loads (as high as many Tbps per connection), satellite and mobile platform movements, switching node states, and atmospheric conditions. This paper presents a multi-layer approach to optical wireless networks and how the architecture can be tuned to specific network constructs and applications.

2. HETEROGENEOUS NETWORK MODALITIES

FSO networks are becoming increasingly important in a wide variety of applications replacing fiber in some cases. Though these communication links are all-optical in nature, there exists a large degree of heterogeneity in their properties and behavior as manifested to the upper layer network protocols. These types of modalities include the following (Fig. 1):

 

Fig. 1. Heterogeneous optical wireless network modalities.

Download Full Size | PDF

A. Physical Layer Communication System Design

For the physical layer and the vacuum links (Type 1), the hard engineering issues have been solved in the past three decades [1–4]. The critical issues in the physical layer are

Figure 2 shows a generic block diagram depicting the above three subsystems among others. This paper is about network architecture rather than the physical layer. For the benefit of being able to understand the context of the rest of this paper, we will summarize the major attributes of a good physical layer design:

 

Fig. 2. Block diagram of an optical space communication system (physical layer only) [2].

Download Full Size | PDF

Figure 3 shows a comparison of the different types of binary signaling schemes and the performance of the corresponding optimum receivers [2]. Modern high-capacity (${\sim}{\rm Tbps}$) optical links require higher dimension signaling to squeeze the modulation into a band-limited receiver detector. A coherent receiver is most effective when there is strong background noise such as bright sunlight. For binary signaling it is almost as good as direct detection, i.e., photon counting receiver. Its capacity for unlimited bandwidth is ln2 bits per photon, the same with intensity modulation and direct detection. However, for large available bandwidth, large symbol size pulse position modulation will yield a capacity of ${{\rm log}_2}{\rm M}$ where M is the symbol size. Its capacity $C$ for bandlimited system is [4]

 

Fig. 3. Optimum receiver performance: probability of detection error for binary signaling—error exponent θ of the tightest exponential bound ${\rm Pr}[\varepsilon] \lt {\rm exp}\{\theta\}$; ${{\rm N}_{\rm s}} = {\rm average}$ number of detected photons per bit [2].

Download Full Size | PDF

 

Fig. 4. Spatial acquisition and tracking geometry [3].

Download Full Size | PDF

Thus, multiple bits per photon capacity can be achieved when background noise is not an issue, and the data rate is modest enough to use M-ary PPM. This is the signaling and detection scheme of choice for planetary explorations. For around Earth operations coherent systems are the choice for their performance in the presence of background noise and phase coherence for multi-spatial-mode coherent detection as treated in a later section. Terrestrial fiber transceivers are being used for space links with a big cost advantage of large-scale commercialization and light weight and small form factor.

All three attributes together will allow the use of the lowest-risk modest power transmitter lasers. Typically, a semiconductor laser with a fiber amplifier will provide enough transmitted power to cover as high as ${\sim}{10}\;{\rm Tbps}$ with a few watts of output optical power. Small-size, low-weight, and power-efficient transceivers have important implications on satellite network architectures. Whereas, previously the optical cross-link hardware has dominated the spacecraft, now more terminals can be flown on the same bus allowing advance network features such as “make-before-break” connections over the network.

B. Satellite Network Physical and Logical Topology Design

Often optical cross-links are used to connect satellite constellations as a “backbone in space” with the uplinks and downlinks being RF or optical, as depicted in Fig. 5 [3]. The constellation can be in geosynchronous Earth orbit (GEO), medium Earth orbit (MEO), or low-Earth orbit (LEO) or satellites in highly elliptical inclined orbits. Figure 6 depicts a notional optical satellite node. Note the possibility of all-optical switching and amplification in the satellite [3], which will be important in some advanced coherent applications for distributed system treated at the end of this paper [9]. Figure 7 is spacecraft node with ${\rm processing} \;+\; {\rm MAC} \;+\; {\rm switching}\;+\;\def\LDeqbreak{}{\rm routing}$ [3]. Not all functions need to be supported.

 

Fig. 5. Optical space backbone for an integrated free-space global network [3].

Download Full Size | PDF

 

Fig. 6. Notional optical satellite node. Note the possibility of all-optical switching and amplification in the satellite [3].

Download Full Size | PDF

 

Fig. 7. Typical spacecraft node with ${\rm processing} \;+\; {\rm MAC} \;+\; {\rm switching}$ and routing [3]. Not all these functions need to be supported.

Download Full Size | PDF

The most important consideration to be made in a satellite network is whether there should be a Layer 3 (routing layer) on the spacecraft, as shown in Fig. 7. Current router designs used in terrestrial networks are power hungry, heavy with large form factors, and not space qualified. Their basic design will not be easily adaptable to space use except for low rates (1–10 Gbps) using hardware and software constructs such as “white box.” It would be a technological tour de force to develop a Tbps router in space. For the foreseeable future (${\lt}{10}\;{\rm years}$), for the backbone in space only Layer 2 switches will be used. With only Layer 2 switching the network will be connection-oriented with ground terminal to ground terminal long-duration tunnels. Any packet switching will have to be done on the ground. This will put severe limitations on the architecture of a fully packet switched space network. For the near term and the rest of this paper we will only consider Layer 2 switched end-to-end connections (${\gt}{1}\;{\rm min}$ durations). This approach will impact the architecture of the space network. Except for geosynchronous networks, all moving satellite networks (with respect to the terrestrial terminals) will have to contend with path rerouting in the middle of a session. Either there is storage somewhere to buffer the data during network topology reconfiguration (hence a delay in data flow of perhaps up to several seconds) or the spacecraft carries extra cross-link hardware to perform “make-before-break” connections. In either case the network management system will have to orchestrate the complicated handover. For a LEO satellite system an end-to-end connection may involve tens of nodes with cross orbital plane links; reconfiguration every few minutes will add a lot of complexity to the path routing management system [10]. The situation with MEOs is much better though not as simple as a GEO network without having to perform path switching in the middle of sessions. Here we will use GEOs as an example to illustrate some of the design considerations involved.

The topology of a satellite network can be viewed in two ways: (1) the physical connection topology with the physical cross-links as arcs of a graph and (2) the logical topology that indicates the path of data flows. Figure 8 shows a GEO constellation with $N$ nodes (satellites) connected physically into a ring network with additional strategically placed cut-through links between satellites to facilitate shorter end-to-end paths for particularly large data flows from sources to sinks (depicted as a circulant graph). The cut-through links of course must avoid the Earth and its atmosphere, but the graphical structures are the same for slightly skewed satellites and cut-through links. The exact logical topology will be highly dependent on the offered traffic pattern. For uniform traffic around the globe, the entrance links and departure links of each satellite have similar loads (uniform, U). For heavy hub traffic (H), the physical logical topology can be asymmetric, as shown in Fig. 8. Figures 8(a) and 8(b) are for uniform offered traffic and Fig. 8(c) is for heavy hub traffic. In the past when CONUS was the source and sink of most of the traffic, the connection (c) was preferred. For the future when data usage becomes much more uniform globally, topologies (a) and (b) should be used.

 

Fig. 8. GEO physical and logical connection topology. The red links are cut-through links to facilitate shorter (reduced number of hops) end-to-end paths for large data flows. (a) and (b) are for uniform offered traffic, and (c) is for heavy hub traffic.

Download Full Size | PDF

 

Fig. 9. Average minimum hop distance with $m$ satellites connected to the source satellite (where the minimum $m$ is 2 because it is a circulant) versus the number of satellites in the constellation [9].

Download Full Size | PDF

 

Fig. 10. Cost of a network with satellites connected to the source satellite versus the number of satellites in the constellation [9].

Download Full Size | PDF

A simple study in [9] was enough to show that the number of GEO satellites should be between 3 and 5 (Figs. 9 and 10) (with the larger number being able to provide back up in case a satellite fails). Also, a couple of cut-through links is sufficient to minimize costs by shortening the number of hops.

A satellite network should never be considered in isolation from terrestrial networks. Fiber networks are ubiquitous and are accessible from many regions of the Earth. Thus, for a satellite network to be viable commercially, it must be integrated as a hybrid network with terrestrial fiber networks; see Fig. 11. Network optimization should be used for satellite network design that can help designers to quantify trade-offs to identify the most cost-effective system for future applications. In [11], a two-stage stochastic programming formulation was used for satellite topology selection, fiber and inter-satellite link (ISL) capacity dimensioning, and routing. This formulation has the advantage of incorporating uncertain traffic demand into network designs and is a powerful tool for capturing the explicit trade-offs between link capacity dimensioning, routing, and overall system cost.

 

Fig. 11. Hybrid paths for data transfer over satellite and fiber networks [11].

Download Full Size | PDF

In [11], we have shown that a hybrid satellite and terrestrial network has a cost advantage over the ISL-only and no-ISL networks. Here we capture the major conclusions of that study:

Figure 12 shows the region of the optimum routing topology selection based on relative fiber and satellite network costs. The ratio of the marginal costs of fiber versus ISLs is the key deciding parameter.

 

Fig. 12. Routing topology selection based on relative fiber and satellite network costs. The ratio of the marginal costs of fiber versus inter-satellite links is the key deciding parameter.

Download Full Size | PDF

C. Satellite Network Critical Technical Issues

Except for the vacuum satellite-to-satellite channel, optical wireless communication systems have attributes not commonly found in wired communications. These are mostly channel effects plus the ability to reconfigure using beam steering. The latter opens up a degree of freedom to rearrange network topologies rapidly on demand, adding tremendous flexibility to the routing layer (Layer 3) that at the moment is not exploited. In addition, there are several areas that are unique to optical wireless network architectures that need attention and unique research and development. Progress in some of these areas is critical to achieve good network performance and, in some cases, vital to acceptable design. Significant technical issues to be addressed for vacuum and atmospheric links and networks include the following (Fig. 13):

 

Fig. 13. Degree of engineering challenges for various links: 1-little, 10-huge. M/P, moon/planets; AC, aircraft; Gd, ground; DC, data center; BH, backhaul.

Download Full Size | PDF

These attributes are not all orthogonal and independent and must be considered jointly during research and system developments. The following architecture areas still need further research:

This paper will suggest a multitude of architecture solutions to the problems. They are reasonable engineering but by no means proven optimum. In some cases, the optimum achievable performance is found analytically, and the solutions typically come close to the limits.

3. OPTICAL SATELLITE AND WIRELESS NETWORK ARCHITECTURE INCLUDING THE TURBULENT ATMOSPHERE

When an optical satellite network connects to an airborne or terrestrial terminal optically, atmospheric and aircraft boundary layer and bow shock turbulence come into play in a major way. Also, the end-to-end path may involve multiple satellite segments and also ground fiber and RF networks as well, as depicted in the dynamic 4-D (four-dimensional, three spatial dimensions plus the time dimension) end-to-end integrated heterogeneous network; see Fig. 14. Due to the wide-ranging behavior of optical wireless networks, manifesting at the link layer as fading, drop-outs due to pointing and tracking errors and long delays, and at the routing layer as fast topology reconfigurations and rerouting for load balancing and congestion control, current network architectures (reflected in its physical layer designs and protocols) will not provide high network performance. To achieve the ultimate potential of FSO, new network architectures are needed in the following directions, as depicted in the protocol stack construct in Fig. 15:

 

Fig. 14. Dynamic 4-D end-to-end integrated heterogeneous network.

Download Full Size | PDF

 

Fig. 15. Protocol stack construct for optical wireless networks.

Download Full Size | PDF

With new architectures, substantial performance gains can be realized in throughputs, delay, reliability/availability, and cost.

A. Physical Layer Challenges

To first order, a Markov channel model captures the statistics of the atmospheric optical channel’s fading process. We model the channel by a two-state continuous-time Markov process, as shown in Fig. 16 [12,13].

 

Fig. 16. Outage in turbulence and two-state Markov mode [12].

Download Full Size | PDF

The states represent the channel as being in an outage (fade) or not. For an optical wireless network, packets transmitted during an outage are assumed to be lost, and packets transmitted during a non-outage are assumed to be received correctly (using an error correction code tuned to the channels statistics). The length of time spent in the non-outage and outage states are exponentially distributed [12]. A typical outage duration ranges from a millisecond to a fraction of a second (${\sim}{1}{-} 100\;{\rm ms}$) with or without diversity. The inter-arrival times of these outages are of the order of 100 ms–1 s. The lengths of time spent in states 1 and 2 (the non-outage and outage lengths, respectively) are exponentially distributed (as a direct consequence of Markov processes and empirically close to measured data [12]). Letting ${Y}$ and ${Z}$ be the outage and non-outage lengths, respectively, their probability density functions are given by

Using the atmospheric optical communication system described in [12], we experimentally justify the exponential distribution modeling of outage and non-outage lengths in moderate turbulence (with the amplitude ${{e}^\chi}$ and log amplitude standard deviation of ${\sigma _{\chi }}\sim{0.1 {-} 0.3}$). In that experiment, samples of the multiplicative power factor (signal strength) were taken every millisecond and for 3 minutes to obtain the histograms in Fig. 17. The histograms are plotted together with an exponential probability density function with the same mean. As seen from these figures, the outage and non-outage lengths are approximately exponential. Whether the channel model is exactly Markov is not very important for the transport layer protocol (Layer 4) architecture, only in that it will provide insightful analytic solutions to the file transfer efficiency if the model is Markov. Since this channel behavior has very bad efficiency for the Internet protocol TCP (transport control protocol), this model is more than sufficient to determine that TCP cannot be used for this channel for big files and a new protocol needs to be created.

 

Fig. 17. Plot of (a) experimental outage lengths (${\rm mean} = {3.69}\;{\rm ms}$) and of the exponential probability density function with the same mean and of (b) experimental non-outage lengths (${\rm mean} = {0.135}\;{\rm s}$) and of the exponential probability density function with the same mean [12].

Download Full Size | PDF

The Markov channel model is further justified when we examined the optical propagation as a higher-frequency extension of the RF case, in which the channel can be modeled as Markov [14] where the spectrum of the log amplitude can be approximated by a one-pole filter. In the optical case, the log-amplitude fading factor has a covariance function $f(t)$ closely approximated by an exponential corresponding to a single pole spectrum; see Fig. 18.

 

Fig. 18. Log-amplitude covariance function—approximate fit to ${{e}^{- ct}}$.

Download Full Size | PDF

B. Diversity Transmission and Reception for Turbulence Mitigation in the Physical and Data Link Control Layers (Layers 1 and 2)

The turbulent atmospheric channel without compensation distorts the phase of the optical field during propagation and results in hot and cold intensity spots at the receiving plane, as shown in Fig. 19 below. If the receiving aperture is in one of the cold spots, fading of the optical signal can be as deep as 10–20 db. The fade duration is of the order of milliseconds and longer and for high-rate channels significant data loss will occur.

 

Fig. 19. Sample intensity distribution of the optical field at the receiving plane without any phase compensation showing hot and cold intensity spots.

Download Full Size | PDF

The use of spatial and temporal diversity is essential at the link layer to combat fading when operating through the atmosphere. In contrast to spatial diversity for wireless systems, spatial diversity for atmospheric optical systems can be readily implemented in a compact fashion since the coherence length is of the order of centimeters; see Fig. 20. The more advanced techniques of using coherent detection with feedback including predistortion transmitter phase arrays (Fig. 21), and a new design for error correcting of block erasures due to fading should be developed. Specifically, the following are high payoff areas of pursuits:

 

Fig. 20. Multi-aperture coherent diversity receiver.

Download Full Size | PDF

 

Fig. 21. Transmission system with dynamically and adaptively reconfigurable receiving plane field patterning via low-rate feedback of the receiving plane optical phase.

Download Full Size | PDF

The major challenges the research in this area must deal with are:

While interleaving plus error correction can be done to recover lost data, interleaving for over a second will present long delays and create significant difficulties for higher layer network protocols. Curiously, if the turbulence is light, the intensity coherence length will be large and for reasonable size telescopes, aperture intensity averaging is not possible. Whereas if the turbulence is strong the coherence length will be short and significant aperture averaging can be used. However, the most effective way of receiver diversity is using distributed apertures, as in Fig. 20.

As alluded to before, spatial diversity for atmospheric optical systems can be readily implemented in a compact fashion since the coherence length is of the order of centimeters (1 cm for strong turbulence and 10–20 cm for moderate and weak turbulence). The available diversity equals the product of the number of independent (in the sense of being in different phase and intensity coherence cells) transmitter elements and receiver elements. Most spatial diversity systems only use receiver diversity and in Fig. 20 [13,15]. Diversity systems can be coherent as in Fig. 20 or a simpler form of incoherent combing with fewer performance gains. An even more effective technique is to sense the receiving plane optical phase and use a low-rate (${\sim}{10}\;{\rm kbps}$) feedback channel between the receiver and the transmitter to implement transmitter phase predistortion to help focus the optical energy on the receiver array, enhancing energy delivery efficiency of the free-space channel; see Fig. 21. Receiver phase tracking element by element either via local oscillator tuning or signal path phase modulation can be used to estimate the atmospheric optical channel state and the data fed back to the transmitter for phase and amplitude modulation of the individual array elements [16,17].

The input/output relationship between the ${n_{\textit{tx}}}$ transmit apertures and ${n_{\textit{rx}}}$ receive apertures is given by a random Green’s function, as shown in Fig. 22. For small discrete apertures such as those assumed here, the input/output field relationships can be summarized by a random matrix ${\boldsymbol H}$ and input and output vectors $x$ and $y$. ${\boldsymbol H}$ has the singular-value decomposition as shown with singular values $\gamma $’s [16,17]:

 

Fig. 22. Random Green’s function relating multiple-apertures’ output field to input field [16,17].

Download Full Size | PDF

Depending on the geometry of the links, either transmitter or receiver or both diversities can be used. For example, for a link from space to ground, receiver aperture diversity can be used whereas it is not available for the uplink since the coherence length at the satellite is larger than the size of the satellite, but transmitter diversity is possible. To use these links in an optical wireless network it is important to characterize the outage probability and outage duration as a function of its design parameters, such as detection scheme (coherent or incoherent) and degree of spatial transmitter and receiver diversity. Outage probability is the probability that the bit error rate of the channel is higher than an outage threshold (where the error correcting code may stop functioning properly). The outage probability and expected outage length for coherence time $t_0$, large degree of diversity $N$, and link margin $m$ is given by [13,15]

For systems with no feedback and predistortion, the outage probability can be significantly reduced with diversity (exponential in the degree of diversity $N$). The expected length of outages decreases inversely as $\surd N$ to $N$, as in Eq. (4). This is very important for the design of the upper layers (Layers 3 and 4) and particularly for the transport layer protocol. Figure 23 shows the performance gains of diversity systems and with feedback over diversity systems without feedback.

 

Fig. 23. Performance gains of coherent diversity systems with feedback over diversity systems without feedback [13,15–17].

Download Full Size | PDF

If the receiver also senses the incoming phase of the optical field and feedback that information to the transmitter, the transmitter can predistort the transmitted field to compensate for the random phase scrambling of the atmosphere. The channel state is an $N$-dimensional ellipsoid with the direction of the major axis as the direction (mode) of the transmitter phases for maximum energy transfer and the eigenvalue is the corresponding power transfer efficiency. Thus, the optimum transmission scheme uses the input eigenvector corresponding to the largest eigenvalue $\gamma^ *$ for maximum power transfer, and the receiver uses the corresponding output vector as a spatial matched filter (Fig. 24) [16,17]:

 

Fig. 24. Channel state as represented by an $N$-dimensional ellipsoid, with tracking via prediction.

Download Full Size | PDF

The feedback channel sends the changes in the $N - {1}$ angles of the major axis of the ellipsoid and the magnitudes of the eigenvalues. When the largest eigenvalue begins to fade below the second largest eigenvalue, the transmitter switches to the new largest eigenvalue and the transmitter mode corresponding to that eigenvalue. Since the turbulence is slow ${\gt}{1}\;{\rm ms}$, the feedback rates are of the order of 10 Kbps that can be transmitted with a low-rate backward link.

A direct detection system with no feedback suffers the classical incoherent combining losses of such a system, ${\sim}$ order of ${N^{1/2}}$. The performance gain of a coherent system is a factor of $N$, which is substantial. By tracking the largest eigenmode of the random transfer function between input and output, not only can the bit error and drop-out rate be reduced but the duration and inter-arrival times of these drop-outs are also reduced [16,17]. The gain of low-rate feedback is another ${\sim}{10}\;{\rm db}$ over that of diversity systems for moderate $N\;{\rm s}$ with no feedback.

C. Routing Based on Prediction, Diversity Combining, Dynamic Route Switching, and Retransmission of Buffered Data—Cognitive Networking

The dynamic physical channel behavior and its impact on upper layer network protocols need a new network architecture and algorithms. It is a perfect setting to apply the new paradigm of cognitive networking, where network intelligence is applied for joint access, routing, and transport layer functions.

1. Cognitive Networking

The bursty and dynamic nature of the traffic generated by future network applications require quick (1 ms–1 s) network adaptation to maintain quality of service and experience. Unfortunately, current network management and control systems are much too slow, and their operational paradigm does not scale well with network size and traffic intensity. A cognitive network management and control system senses current network state conditions such as traffic and flow patterns and uses this information to decide how to adapt the network to satisfy/improve overall performance and provide quick responses to transaction requests. The cognitive network module is part of the control plane that touches all layers of a network; see Fig. 25 (it may reside at all the network nodes as well as at a centralized or distributed controller/s). The cognitive network management module is a collection of coordinated algorithms that sense and infer network states; decide and implement fast scheduling of flows; perform rapid load balancing; and handle rapid reconfiguration, restoration, reconstitution of diminished, compromised or failed network assets. It can also predict evolution of network states and intention of users and take appropriate actions. However, one will be faced with a network whose links are dynamically changing perhaps as fast as milliseconds. The number of network parameters and their short coherence time for the very dynamic optical wireless networks render acquiring the complete state information of the network impractical. The idea is to sense and control a small subset of the parameters with the “largest” information contents and use and their relationship to end-user performance to maximize performance [18,19]. Cognitive techniques that need to be incorporated in the strategy include

 

Fig. 25. Cognitive network management and control system, touches all layers of a network.

Download Full Size | PDF

Cognitive techniques come in many forms. Several techniques are used currently such as probabilistic Bayesian-based inference, machine learning, graph analytics, and neuromorphic computing. In this paper we use a combination of Bayesian probabilistic and learning techniques to illustrate the cognitive network control technique for optical satellite networks.

2. Control Plane Link State Sensing and Link Establishment (Data Link Control, Layer 2)

It is possible to dynamically configure the nodes of an optical wireless network to provide any desirable physical connection; see Fig. 26. Since there is no fiber and only free space, there are no connectivity constraints. With beam switching speeds of ${\lt}{1}\;{\rm ms}$ within a node and a telescope slew rate of $\sim{\rm seconds}$, the reconfiguration can be essentially as fast as it needs to be. Buffering and synchronization at nodes is needed for packets and for all-optical switching buffering is at the source.

 

Fig. 26. Graphical model of optical wireless network nodes capable of fast rerouting and beam steering.

Download Full Size | PDF

We assume the existence of a low-rate but reliable control network that has electronic routing and switching. The cognitive networking module/s is/are attached to the control network and is/are both centralized and distributed depending on the function. This control plane shares information on the traffic and the channel states of all the nodes. The control plane touches at least four layers of the network from Layer 1 to Layer 4 and in some cases the application layer. The network architecture can have one of several modes of transport:

Each node ${{\boldsymbol N}_i}$ is assumed to keep track of the state of its connection ${l_{\textit{ij}}}$ to node ${N_j}$; see Fig. 26. Specifically, it will note the largest eigenvalue of the connection and the next several largest eigenvalues with the possibility of reaching node ${{\boldsymbol N}_j}$ via another connection. It can use the best of the paths for transmission or use several for diversity transmission or as hot spares. For example, the node can use the following three paths, two from the same telescope connection and the third from another path with non-overlapping links:

The estimation problem relies on the feedback from the receiver, which is slightly stale due to estimation, compression, and feedback delays. The channel is approximately log-normal and estimators for Gaussian channels can be used. If there are ${\boldsymbol N}$ nodes in the network, we assume we know the mid-term average capacities of the ${\boldsymbol N}({\boldsymbol N} - 1)$ pairwise links. This can be in matrix form, $C = \{{c_{\textit{ij}}}\}$, and can be used in an algorithm for topology reconfiguration.

Except for short drop-outs (${\sim}{50}\;\unicode{x00B5} {\rm s}$), the node can predict ahead of time when the predistortion spatial pattern corresponding to the largest eigenvalue yield to the pattern with the second largest eigenvalue. Buffering at each node will retransmit faded data via an automatic repeat request (ARQ) in Layer 2. Data loss due to short fades can also be recovered via interleaving and error correction or in the case of path diversity transmission the use of an erasure code across paths.

In the all-optical flow case, the lack of an all-optical memory necessitates a different treatment. Either the sender needs to recognize a degradation of the flow path or a backward control message from the receiver to Layer 4 of the transmitter will inhibit flows until after successful reconfigurations. An alternative to deal with topology reconfiguration is to have a spare aperture and use “make-before-break” to establish a new connection before dropping a declining link, but this technique still needs the sender to temporarily stop sending data to prevent data overlap in case the new path is shorter timewise than the previous path.

3. Traffic Matrix Sensing, Estimation, and Prediction

In most cases, the network offered traffic is typically bursty and unscheduled. Thus, the offered traffic needs to be sensed and the state of congestion at each node must enter into a path assignment algorithm. We concentrate on the arrival of large elephant flows at the nodes, since they make up most of the traffic volume and cause large surges in delays. For every node $j$ we can estimate the arrival rate of large flows to node $k$, by considering the arrival times of the flows $\underline{t} = \{{t_i},(- \infty ,t)\}$. The minimum mean square estimator of the arrival rate $\lambda$ is  [20]

The value of $T$ and $n$ must be chosen to be sufficiently large, so the estimation error is kept small. A small (normalized by the mean $\mathfrak{m}$) standard deviation $\sigma /\mathfrak{m}$ can be used with the Chebyshev inequality to guarantee small false alarm rates and high probability of detection in a Neyman–Pearson type test.

Often for real traffic the prior does not have an analytical form and the actual density must be acquired via learning of historical data. The first step of the estimation algorithm is to try to form the estimator of the traffic matrix $T$ by multiplying the estimated arrival rate $\lambda$ by the average session size, ${E\{S\}}$:

4. Physical Topology Initiation and Reconfiguration

The optical satellite network needs to initiate a physical connection topology [18]. On cold start the network will form the normalized traffic matrix $B$ by dividing $T$ by $C$ element by element as shown below. The number of nodes any given node can connect to is limited by the number of telescopes it has. Note the satellite or terrestrial node can have more transceivers than telescopes because of the possibility of using wavelength-division multiplexing or spatial mode diversity with the same phase array transmitter. If a phase array transmitter is not used, then spatial mode diversity with feedback as described cannot be used:

Given the matrix $B$, the optimum algorithm to find the connection topology $G^*$ is a mixed integer non-linear (convex) program which is hard to solve. If desired, the optimum problem can be solved off-line on initiation but on-line adaptation will be computationally challenging. However, the following greedy suboptimum algorithm is excellent and suffices in most cases:

To deal with the evolving time series of the problem we can also use a time sequential test first proposed by Abraham Wald, called stopping trials [21,22]. A “stopping trial” function $J$ is used to trigger reconfiguration using a maximum likelihood, MAP rule, where the log-likelihood functions ${L_{\textit{ij}}}(t)$ of the estimators of $B$ are used to pick ${\Delta _{\textit{ij}}}$ nodes (which equal to the number of transceivers and telescopes at each node), with largest values to connect to, generating the connection graph ${ G}(t)$. The function $J$ can be chosen to be the load of the link, $\rho$, or the delay, $d$, due to the congestion on the link using a queueing model. Thus, $J$ can be a non-linear but a one-to-one function of the underlying arrival process of the flows. Such an algorithm is very responsive to temporary node failure or impairments, e.g., tracking system glitches, single event upset in node hardware/software, and deep fading causing loss of tracking. It is probably the fastest sequential on-line algorithm (in the statistical sense) to perform reconfiguration. The only assumption needed is that the traffic arrivals are independent [21,22]. As alluded to before, proper buffering or backward inhibition of the transmitter temporary from transmitting is necessary to make sure flows do not overlap or data is lost during reconfigurations. This can only be implemented when the control plane touches Layer 4 at the users’ interface and at intermediate nodes with storage.

For GEO satellite networks, the physical connection topology is trivial, and the optimum solution can easily be found without resorting to the suboptimum algorithm. Reconfiguration will also be infrequent. For LEO constellations with thousands of satellites, the problem is extremely complicated where in-plane satellites and cross-plane satellites have to be treated differently because of their durations of reconfiguration times. Thus, for LEOs, cross-plane connections should be discouraged by putting the appropriate weight (penalty) on such connections.

5. Routing

Due to the weight, power, and form factor of current generations of routers, it is unlikely a satellite network will have a full-rate (i.e., enough capacity to allow packet switching of all traffic) router on each satellite. They may have low-rate “white box” routers for low-volume traffic that will be a small fraction of the traffic. Most of the traffic will be connection-oriented end-to-end sessions. This is actually good for the routing problem where it will be unnecessary to compute the routes packet by packet but only the route at session initiation. This path will maintain until the session is over or when a reconfiguration is required as in setting cross-plane LEO connections. Most of the routing problems can be solved by standard routing algorithms such as Dijkstra’s algorithm. The weights for each link can be adapted to the satellite system’s specific cost drivers and performance metrics. In most cases the overall end-to-end cost is additive, lending themselves to optimum routing via dynamic programming algorithms such as Dijkstra’s algorithm.

While the routing weights (cost function) of GEO and MEO networks can be the usual forms such as using the load of the link as weights and minimizing the overall congestions and/or adding to the weight a term that represents propagation delay to discourage long circuitous paths. For LEO constellations the problem is very complex due to the need for frequent reconfigurations with satellites rising and setting and also short cross-plane contact durations. A recent analysis of a typical LEO constellation revealed that the routes that minimize the number of hops, minimize cross-plane connections, and maximize connection durations are very different [10]. Dijkstra’s algorithm works in that case but the cost function $\omega_{ij}$ must have at least three terms:

The above discussions only cover the vacuum channel part of the networks. When the connections have links that go through the atmosphere or aircraft boundary layers, the fading turbulent channel effect must be accounted for in the routing algorithm. We will use a simple homogenous model to illustrate how this can be done. Extensions to networks with mixed vacuum and atmospheric links will be simple.

For simplicity of illustration, assume all links have turbulence where the feedback channels provide perfect channel states so the transmitter will predistort the transmitted phases to use the eigenmodes corresponding to the largest eigenvalues for maximum signal power transfer to the receivers. Also assume hard decisions on each symbol received are made at each node and error correcting code is used only end-to-end, i.e., without decoding and recoding at intermediate nodes. With the proper scaling the error probability at each node is of the form

This assumes optimum demodulation of binary or M-ary signaling and ${c_l}$ is a constant that captures the power transfer efficiency of the telescopes for that particular link $l$ as well as the modulation/demodulation performance. For a path $P = \{l\}$ with a set of links $\{l\}$, the end-to-end probability of error is

6. Diversity Routing for Reliability

For satellite networking paths with vacuum links only, the only source of errors should be from detection noise and background noise. Beam tracking should be working well so that no power fades due to beam tracking error should be present. In the case where turbulence is a factor, even with feedback and transmitter phase predistortion, there will be residual burst errors caused by unmitigated fades. The cause of these fades, albeit rare and short, can be the result of phase estimation errors at the receiver, stale feedback data due to processing and propagation delays, and imperfect phase modulation at the transmitter and receiver. When multi-path routing is possible in a multiply connected network, each path will fade independently, with the probability of all paths fading at the same time being very small though not zero. In this case, diversity routing can be used to improve reliability [23].

To find the diversity routes, first find the route with the best performance as before. Then delete the most vulnerable links in sequence from the network graph and rerun the routing algorithms to find the next link disjoint path. For $n$ paths pick any (${\rm L},{\rm K}$) erasure code of length ${\rm L}$. If ${\rm K} \lt {\rm L}$ bits is transmitted successfully and not erased, the code can correct the message. If ${p_i}$ is the probability of error due to fades of path $i$, the optimum allocation of the ${\rm L}$ bits to link $i$ is ${{\rm L}_i}{p_i} =$ constant and ${\rm L} = \Sigma {{\rm L}_i}$. This allocation of coded bits per link is such that the expected number of successfully transmitted bits/path are equal though they have in general unequal reliability. However, this scheme has to pick ${\rm L}$ and ${\rm K}$ before transmission and residual errors can still exist. As in common networking practice, these residual errors or errors when diversity routing is not used should be taken care of by the transport layer protocol in Layer 4.

4. TRANSPORT LAYER PROTOCOL

Currently the TCPs and variants for large bandwidth-delay satellite channels are not adequate of the high-rate turbulent atmospheric channels. In this section we will address the reasons why a new Layer 4 protocol other than TCP needs to be constructed.

A. Transport Control Protocol

The TCP (currently used in Internet at Layer 4) is used for reliable data transfer in a packet switched network. For the Internet it performs four functions:

Since we assume for the near future there is no router in the satellites and the transport is connection-oriented, the first two functions are not needed. In fact, the second function creates serious inefficiencies for large bandwidth-delay product channels such as an optical satellite channel. There are stop gap transport layer protocols developed for RF satellite channels, but they are not good enough for the even higher-rate optical channels with addition of fades due to turbulence. The state of a TCP session can be represented by the window size of the number of packets that can be sent before acknowledgement. The upper limit of the window is usually set at 128 packets for fair capacity allocation. However, due to the large bandwidth-delay product of a satellite channel, this upper limit is usually lifted to allow good throughput albeit at the expense of fairness. To deal with congestion in the Internet, TCP works with routers to use window closing for congestion control. When a router is congested it will discard packets via buffer overflow and TCP will react by closing its transmission window from its maximum allocated window size. For single packet losses, TCP will go into “fast recovery” mode halving its window size and only build the size up linearly after each successful round trip without packet loss. However, for fading channels this mechanism is seldom triggered. When a fade erases many packets in sequence, TCP timeouts and it will initiate its buildup from one packet in flight as if it is going through a new start; see Fig. 27 [24]. The saw tooth behavior of buildup and slow start will limit the throughput of TCP even with the upper limit of the window size is lifted; see Fig. 28.

 

Fig. 27. Discrete time Markov chain model of TCP’s flow control when outages cause timeouts and “slow start” [24]. The states of the chain represent the window size of the number of packets inflight without acknowledgement, and ${{n}_{\rm{max}}}$ is the upper limit of the window size. $p$ is the probability of successful transmission [24].

Download Full Size | PDF

 

Fig. 28. Atmospheric link TCP throughput efficiency (with zero congestion loss) versus round-trip distance at 10 Gbps and 8 db of link margin under moderate to strong turbulence [25].

Download Full Size | PDF

Figure 28 shows the efficiency of TCP versus round-trip distance (propagation delay only) assuming zero processing delay. Only fading drop-outs are present with no other detection errors. Even with LEOs, the protocol has poor efficiencies [25].

An attempt to quantify what diversity transmission might help to increase reliability [26] explored an extreme form by repeating each packet $n$ times over $n$ paths. Thus, the efficiency is limited to ${1/}n$. Figure 29 shows the TCP efficiency for an extremely turbulent channel when the inter-arrival time of fade is 1 s and a fade duration of 1 s. Repeat code is used because it is asymptotically optimum for extremely poor channels. The efficiency has improved some over TCP with no diversity but is still not sufficient.

 

Fig. 29. TCP efficiency in multi-path diversity with repeat coding with rate 1 Gbps, packet size 10 kb, round-trip time of 0.1 s, TCP maximum window size of 1000/round-trip time, expected fade inter-arrival time of 1 s, and expected fade duration of 1 s [26].

Download Full Size | PDF

B. New Transport Layer Protocol for Flows over an Atmospheric Channel

The dynamic link states present wildly varying qualities of packets delivered to Layer 4. For an efficient design, Layer 4 must be designed with Layers 3, 2, and 1 designs in concert. The design requires an active control plane with fast network state sensing. The control plane senses the frequency of drop-outs due to fading. Rate matching can be done by prior agreement for big flows. The frame size of a big flow is chosen so that the error correcting code can efficiently correct the number $\Gamma$ of burst erasures due to the outages; see Fig. 31 below. Frames with uncorrectable errors are recovered via ARQ; see Fig. 30. Entire uncorrectable frames are retransmitted at the interlaced select repeat duration (Fig. 30). Figures 31 and 32 both show the performance of this new protocol, called HH-L4 [25], as compared to TCP. Both the throughput efficiency and resulting transfer delay are substantially improved with close to 100% efficiencies.

 

Fig. 30. Flows segmented into large frames with ARQ select repeat frames interlaced after frames [25].

Download Full Size | PDF

 

Fig. 31. Throughput efficiency of long frames in the presence of burst errors with optimum and practical frame size as a function of the number of correctable drop-outs $\Gamma$ per frame via frame ARQ [25].

Download Full Size | PDF

 

Fig. 32. Comparison of throughputs of TCP and HH-L4 of a 10 Gbps path [25].

Download Full Size | PDF

The number of fades $\Gamma$ in a frame must be sensed by the network management and control system and the frame size selected accordingly. Frames sizes are not very sensitive to the state of the turbulence and can be chosen as larger than the optimum size with little sacrifice in performance; see Fig. 31. Typical sizes are between 100 Mb and 1 Gb, in contrast to the IP packet size of ${\sim}{10}\;{\rm kb}$.

Figure 32 shows the normalized delay of this protocol as compared to TCP [25]. The delay is normalized to one transmission time of the file. This protocol exhibits low delay because it is designed with burst errors and big flows in mind.

5. FUTURISTIC APPLICATIONS OF A HIGH-SPEED OPTICAL SATELLITE NETWORK

With coherent communication between satellites and airborne and terrestrial nodes and the possibility of all-optical routing and switching as in Fig. 6 (transport mode 2b in Section 3.C.1), the optical satellite network can become a distributed sensing and communication system as depicted in Fig. 33. Optical amplifiers at each node can boost the signal energy before retransmission.

 

Fig. 33. Distributed networked satellite systems.

Download Full Size | PDF

The coherency down to the optical wavelength level will allow the realization of a number of futuristic applications. In addition, the space backbone allows the presence of a shared spaceborne processing satellite that is made up of a collection of commodity processors as in a terrestrial computing cloud. Radiation hardening is not an issue because these processors need to live for only 3–4 years with failed elements replaced by other heavy elements via scheduling. Spaceborne processing before downlink will alleviate the volume of raw data being downlinked and allow better coherency among distributed satellite platforms. The following is a list of possible new applications [3,9]:

Since the satellite hardware cannot be replaced easily once launched, it is imperative some foresights be used to configure the satellite hardware for such futuristic applications. The processing and algorithm aspects can be modified as needed with replaceable processing satellites and unloadable software.

6. CONCLUSIONS

For optical satellite networks, the physical and data link control layers are mature. The hard problem of pointing and tracking of narrow beams has been solved over the past two decades. The use of fiber optics commodity coherent transceivers has lowered the cost, weight, and power consumption substantially. Thus, there is a proliferation of satellite networks deployed or under construction. Underdeveloped are the routing and transport layers protocols. A well-designed satellite network will need much better coordination among the network layers. Some of the deployed systems may or may not have the necessary hardware to support the best upper layer protocols. For example, for LEO systems extra cut-through cross-links are likely required to reduce hop counts and processing delay for long-distance connections. Some of the already launched satellites may not have enough telescope to support a more dynamic and adaptive physical topology. Finally, the optical space network is not a direct replacement of terrestrial fiber and wireless networks. It can be used as an augmentation, but it also offers many more avenues for applications that terrestrial fiber and wireless networks cannot reach. The creativity of new services is still very much ahead of us.