Citation Context

...lem is to process the data closer to the sensor that generated them, thereby reducing the need for backhaul communications. To achieve this, parallel data processing infrastructures such as MapReduce =-=[5]-=- could be used in the field as a way to handle the data volume while still meeting short deadlines. Of course, delivering the necessary computing power to sensors in the field is not an easy task, esp...

Citation Context

... with H at a point near the axes, where the sparse vector lies. Fortunately, the ℓ1 minimization problem is tractable; it is a convex optimization and reduces to a linear program called basis pursuit =-=[44]-=-, which is solved by interior point methods. A full treatment of the optimization methods used to decode is beyond the scope of this chapter and the interested reader is referred to the ℓ1-magic packa...

Citation Context

...or decode, the original signal from y given Φ and Ψ, even though solving for s is an ill-posed problem since M < N? Below, we consider these two issues in turn. 2.4.2 Encoding Candés, Romberg and Tao =-=[39]-=- have identified two necessary and sufficient conditions under which constructing Φ will permit accurate recovery of the original signal. as The first condition is the incoherence of Φ with respect to...

Citation Context

...lector’s Problem and packet erasures). We will describe these gains and the mechanics of network coding in detail below. 2.3.1 Throughput Gains In the seminal paper on network coding, Ahlswede et al. =-=[25]-=- proved the theoretical bounds on the capacity of a multicast network and showed that traditional multicast routing could not achieve the bound, but that network coding could. Consider the butterfly n...

Citation Context

... 2 In this example, the input key is irrelevant.Chapter 2: Background 29 will assume here since a full treatment is beyond our scope. The interested reader is referred to the work of Ghemawat et al. =-=[21]-=- for details. A MapReduce job, which consists of M map tasks and R reduce tasks, is administered by the MapReduce run-time according to the following execution flow. First, the user submits her job to...

Citation Context

...on process. Below, we will discuss the basic formulation of the high-rate sensing problem and the mechanics of compressive sensing encoding and decoding, as they are typically found in the literature =-=[41, 42]-=-. Later, in Chapter 5 we will explore how CS is a natural fit for the distributed source setting. 2.4.1 The High-Rate Sensing Problem Consider a real-valued, finite-length, discrete-time signal repres...

Citation Context

...s that if s were approximately K-sparse, then our recovered ˆs signal is almost as good as if we had known the indices of the K largest coefficients in s and just measured these coefficients directly =-=[40]-=-. Not only does this allow us to handle noisy signals, but this property also lays the foundation for the ability to perform incremental decoding of compressive measurements in a “largest coefficient ...

Citation Context

...satisfied through in-network aggregation of raw sensor data as the data traveled up the collection tree, and offered a simple form of data compression. Among these works were the Cougar [100] and TAG =-=[101]-=- projects, which treated a sensor network’s data streams as an append-only relational database, against which SQL-like queries using aggregation functions (such as MIN, MAX, COUNT, SUM and AVERAGE) co...

Citation Context

... X had access to “side-information” Y , then X and Y could coordinate their coding and use the optimal total rate of RX + RY ≥ H(X, Y ), the rate of a joint optimal encoder. In 1973, Slepian and Wolf =-=[108]-=- published a landmark paper proving their eponymous theorem stating that X and Y could be coded with a total rate equal to the joint entropy H(X, Y ) as long as their individual rates satisfied RX ≥ H...

Citation Context

...d (called the native data) are treated as a vector over a finite field. A linear transformation is then applied to the vector (i.e., a linear combination is taken) and the result forwarded. Li et al. =-=[27]-=- showed that, for acyclic multicast networks, linear network codes achieve the maximum multicast capacity bounds; Koetter and Médard [28] extended this result to arbitrary networks. Subsequently, Ho e...

Citation Context

...yk,1 . . . yk,n 0 0 . . . 1 xk,1 . . . xk,n In this form, the native packets x1, . . . , xk can be read out from the coding matrix easily. 2.4 Compressive Sensing The Nyquist-Shannon sampling theorem =-=[37]-=- states that a signal must be sampled at a rate that is at least twice the maximum frequency present in the signal (i.e., at the so-called Nyquist rate) in order for signal reconstruction to suffer no...

Citation Context

...near combination is taken) and the result forwarded. Li et al. [27] showed that, for acyclic multicast networks, linear network codes achieve the maximum multicast capacity bounds; Koetter and Médard =-=[28]-=- extended this result to arbitrary networks. Subsequently, Ho et al. [29,30] showed that the above properties hold when random coefficients are chosen for the linear transform matrix, i.e., when each ...

Citation Context

...ion details. It is, however, worth noting here that the bulk of these numerical methods has a non-trivial observed polynomial complexity of at least O(N 3 ), though more recent methods such as CoSaMP =-=[46]-=- achieve O(N 2 ) with a slight increase in the number of measurements required. Still, there is an explicit trade-off when using compressive sensing: simplicity during encoding is traded for complexit...

Citation Context

...centage of the total population, the signal is sparse. Moreover, the signal tail drops rapidly, meaning the number of measurements required to accurately decode the largest magnitude anomalies is low =-=[97]-=- and decoding can occur early, as soon as the first few measurements are received. In Evaluations 5.3.2 and 5.3.3 below, we are interested in knowing the effect of M on decoding accuracy for such sign...

Citation Context

...or different neighbors and broadcasting the coded packet. Loss correction is achieved via hop-by-hop ACK and an external best-path routing mechanism is assumed. Results from deployment on MIT RoofNet =-=[83]-=- showed that COPE could increase the throughput of TCP flows, though this is achieved by making COPE TCP-aware to prevent TCP packet reordering. Instead of working on multiple flows, the MORE project ...

Citation Context

... destination to source, can be similarly encoded but in a different generation ID space. Thus, FlowCode does not seek the traditional network coding throughput gain due to intersecting packet streams =-=[57]-=-. We have found that behaviors of complex transport protocols like TCP (e.g., delayed ACK) present few opportunities to achieve this type of gain in practice. Furthermore, for the TCP ACK stream flowi...

Citation Context

...omputing power to handle decoding separately. This is one of the most attractive advantages to compressive sensing and has been exploited already in many sensing modalities, including medical imaging =-=[47]-=- and radar [48]. Yet another advantage of compressive sensing is its robustness against signalsChapter 2: Background 49 that are not exactly sparse. In our discussion so far, we’ve only considered K-...

Citation Context

... can provide additional gains in more sophisticated settings than the canonical examples above. 2.3.2 Robustness Gains Random linear network codes provide robustness to the Coupon Collector’s Problem =-=[31]-=- as well as packet erasures. These gains are particularly important to practical systems that implement network coding, since they generally help simplify protocol designs, as we shall see below.Chap...

Citation Context

... queries were satisfied through in-network aggregation of raw sensor data as the data traveled up the collection tree, and offered a simple form of data compression. Among these works were the Cougar =-=[100]-=- and TAG [101] projects, which treated a sensor network’s data streams as an append-only relational database, against which SQL-like queries using aggregation functions (such as MIN, MAX, COUNT, SUM a...

Citation Context

... (e.g., via feedback), then the transmitted signal x can be recovered by channel matrix inversion. MU-MIMO has been recently adopted in various revisions to broadband wireless standards such as WiMax =-=[68]-=- and 3GPP LTE [69], though base stations and devicesChapter 4: Improved Data Transport Using Network Coding 118 that support these extensions are not yet widely available. Base stations that are comm...

Citation Context

...ibe in detail how random linear network codes work in practice. In particular, we describe the implementation of random linear network coding we will use in Chapter 4, which is similar to other works =-=[35, 36]-=-. We assume an arbitrary multihop network in which one node is a source and another a destination. The source generates an infinite stream of native packets destined for the destination. We assume nat...

Citation Context

...ns that data reduction solutions based on local comparisons, filtering, and summarization are unsuitable. Existing solutions resort to reducing the data volume by either employing aggregation methods =-=[87]-=- to lower the resolution of information or by sampling at a low rate [88]. Unfortunately, neither strategy is amenable to the continuous, fine-grain monitoring we want. This led us to cast the scenari...

Citation Context

...ed using just 2 bits, which is the same rate achieved as when the encoder had access to side information V . More than 30 years after the Slepian-Wolf theorem, Pradhan and Ramchandran proposed DISCUS =-=[109]-=-, one of the first practical methods for partitioning the input space into co-sets according to the channel code being used. However, such distributed source codes did not support different compressio...

Citation Context

...ata and routing were jointly optimized. Initial simple models were route-centric, where data were optimally routed via a shortest path tree and only opportunistically compressed at intermediate nodes =-=[104]-=-. These evolved into more data-centric approaches where sub-optimal routes were used to maximize the correlation between data arriving at relays. Scaglione et al. proposed a heuristic for determining ...

Citation Context

...d that, for acyclic multicast networks, linear network codes achieve the maximum multicast capacity bounds; Koetter and Médard [28] extended this result to arbitrary networks. Subsequently, Ho et al. =-=[29,30]-=- showed that the above properties hold when random coefficients are chosen for the linear transform matrix, i.e., when each node uses a random linear network code to generate random linear combination...

Citation Context

...ength (RSS), even if the difference in RSS is small. This phenomenon is known as the capture effect and has been known to occur in a wide variety of transceivers besides 802.11, such as ALOHA network =-=[16]-=-, FM [17], and Bluetooth [18]. The difference in RSS between two competing transmitter’s frames can be viewed as a signal-to-noise ratio (SNR), with the stronger transmission considered the signal and...

Citation Context

...ransport. For example, it is well-known that under these lossy conditions, transport protocols such as TCP achieve considerably lower throughput than that offered by the capacity of the wireless link =-=[51]-=-. This is because TCP misinterprets packet loss as a congestion signal, resulting in unnecessary source throttling. Furthermore, during long-term outages, TCP will timeout and go into exponential back...

Citation Context

...nd summarization are unsuitable. Existing solutions resort to reducing the data volume by either employing aggregation methods [87] to lower the resolution of information or by sampling at a low rate =-=[88]-=-. Unfortunately, neither strategy is amenable to the continuous, fine-grain monitoring we want. This led us to cast the scenario as a distributed compression problem (e.g., nodes with “normal” status ...

Citation Context

...creasing the time it takes to establish a connection if k were large. Throughput gain from inter-flow network coding is negligible in bulk data transport. Unlike other wireless network coding systems =-=[57,63]-=-, FlowCode does not attempt to achieve the throughput gain resulting from coding together independent packet streams that intersect at a network node (i.e., inter-flow network coding). Net-Chapter 4:...

Citation Context

...tion domains, ranging from indexing for web search [5] to scientific data analysis for high-energy particle physics [22] and computational chemistry [23] to machine learning in recommendation systems =-=[24]-=- and image analysis [7]. Furthermore, the traditional MapReduce deployment onto large commodity clusters has made big data processing economical for even small organizations, since the hardware is bot...

Citation Context

...antenna patterns and the capture effect. We base our discussion of path loss and fading on the treatment of the subjects by Rappaport [13] and Parsons [14], and of antenna patterns on that by Balanis =-=[15]-=-. The interested reader is referred to these texts for further details. 2.1.1 Path Loss As a receiver moves away from a transmitter, the power of the signal it receives will drop, since the transmitte...

Citation Context

...ormation is therefore crucial for better decision-making by automated data center management systems [85] and for maintaining what amounts to a geographically-distributed, container-based data center =-=[86]-=-.Chapter 5: Low-Overhead Status Monitoring using Compressive Sensing 131 Second, as UAVs are recruited by a terrestrial unit, a typical operating scenario is one in which the ground unit assigns task...

Citation Context

...re suited to handle data with low and high correlations, respectively. Neither perform particularly well when there is an intermediate level of correlation. To address this shortcoming, Pattem et al. =-=[105]-=- proposed a hybrid approach that exploits the fact that spatially close sensors are likely to generate highly correlated data. In their approach, the sensor network is first grouped into geographic cl...

Citation Context

...s led us to cast the scenario as a distributed compression problem (e.g., nodes with “normal” status are essentially correlated sources, in the parlance of distributed compression for sensor networks =-=[89]-=-), for which a solution naturally arises: in-network compression of status messages using compressive sensing. We observe that the required bandwidth for each network link in a status collectionChapt...

Citation Context

...d that, for acyclic multicast networks, linear network codes achieve the maximum multicast capacity bounds; Koetter and Médard [28] extended this result to arbitrary networks. Subsequently, Ho et al. =-=[29,30]-=- showed that the above properties hold when random coefficients are chosen for the linear transform matrix, i.e., when each node uses a random linear network code to generate random linear combination...

Citation Context

...sing chunks of a file it has, it codes all of its chunks together and broadcasts the coded chunk to help fill gaps at all of its neighbors. Lee et al. [78] formalized this idea into a BitTorrent-like =-=[79]-=- protocol called CodeTorrent for file sharing in simulated vehicular networks. In both these systems, nodes promiscuously listen for coded packets from neighbors. Naturally, efforts in file sharing an...

Citation Context

...t the transmission operation that node Z performs is a broadcast of b1 ⊕ b2 to the destinations. This suggests that network coding is particularly well-suited to wireless, which is a broadcast medium =-=[26]-=-. We illustrate how network coding can improve the throughput of a wireless network in Figure 2.5. Consider wireless node A with packet a and node B with packet b. They wish to exchange packets throug...

Citation Context

...e receiving a value of “1” from each of its two children would transmit a value of “2” to its parent, thereby reducing the number of transmissions by 50%. This work was extended by Shrivastava et al. =-=[102]-=-, who introduced the quantile digest, a data structure that adapts its compression to the data distribution by binning the data into variable-sized buckets of approximately equal weight. This data str...

Citation Context

...ks are fixed. In the wireless sensor network setting, variable-rate compression is necessary because the data correlation, and thus the compression rate, can be time-varying. In response, Chou et al. =-=[110]-=- propose a practical, two-phased approach for distributed source coding that supports variable-rate compression without the need to change the underlying codebook. In Phase 1, the decoder gathers unco...

Citation Context

...essage to a randomly chosen node out of the other n−1 nodes— would require O(n log n) rounds. However, if the nodes performed random linear network coding, the number of rounds required would be O(n) =-=[32]-=-. This is again due to the fact that each RLC received is as useful as the next, so each node needs to collect just n RLCs. Packet Erasures It is generally difficult to achieve simultaneously the opti...

Citation Context

...uable systemlevel insights. Generation size should be small. There are several practical reasons for keeping the generation size small, even though larger generation sizes are traditionally preferred =-=[35, 62]-=- since they defend more effectively against the coupon collector’s problem. Primarily, FlowCode must keep generation sizes small because it should not introduce extraneous burstiness to transport laye...

Citation Context

... rounds, after which Phase 1 is re-invoked to limit error accumulation. An interesting consequence from Slepian-Wolf coding is that routing can be completely decoupled from the problem of compression =-=[111]-=-. Thus, it circumvents the NP-hard problem of jointly optimizing rate allocation and routing: the optimal route is the shortest-path tree and the optimal rate allocation is given by Slepian-Wolf codin...

Citation Context

... the maximum number of concurrent connections. We call this modification TCP with low-rate retry (TCP+LRR). Our validation experiments compare the performance of unmodified, standard Linux TCP (CUBIC =-=[58]-=-) over FlowCode against that of TCP+LRR over the standard 802.11 MAC with link-layer ARQ enabled. For TCP over FlowCode, we use parallel links in a 4-Tx/4-Rx configuration (as described in Chapter 3) ...

Citation Context

... own SNR thresholds for capture. For 802.11 radios, the capture threshold can be <1dB, but is typically a function of the PHY bit rate, with higher bit rates requiring a higher threshold. Recent work =-=[19,20]-=- in 802.11 networks have shown that capture occurs under three distinct scenarios and that some of these cases have different SNR thresholds. For simplicity, we describe these scenarios using only two...

Citation Context

...e it is essentially white noise, an i.i.d. Gaussian Φ is incoherent with respect to almost any basis Ψ. Other random matrices, such as those with entries drawn from a symmetric Bernoulli distribution =-=[43]-=-, have also been shown to satisfy the incoherence and RIP conditions.Chapter 2: Background 46 ˆs ˆs s H s H ˆs s H (a) ℓ2-norm minimization (b) ℓ0-norm minimization (c) ℓ1-norm minimization Figure 2....

Citation Context

...lt to achieve simultaneously the optimal delay and optimal rate in a packet erasure network using existing methods such as link-by-link block coding or end-to-end erasure codes such as fountain codes =-=[33]-=-. Typically, the former trades delay for a higher rate, while the latter trades rate for a shorter delay. Consider a simple linear network topology with three nodes A, B, and C. Link AB has a loss rat...

Citation Context

...e.g., multicore processors and GPUs) as well as its wide use in a variety of application domains, ranging from indexing for web search [5] to scientific data analysis for high-energy particle physics =-=[22]-=- and computational chemistry [23] to machine learning in recommendation systems [24] and image analysis [7]. Furthermore, the traditional MapReduce deployment onto large commodity clusters has made bi...

Citation Context

...ering in vehicular 802.11 networks and multi-user MIMO (MU-MIMO). We will first consider these separately and then discuss some recent work on UAVs in both areas. Beam Steering. The MobiSteer project =-=[64]-=- investigated the use of beam steering based on probing to improve the throughput of 802.11 links between a moving car and roadside access points (APs). MobiSteer uses a car-mounted, software-controll...

Citation Context

...and (3) it was designed to process large volumes of data effectively and has been used effectively in application domains of interest (e.g., image clustering [7] and, more generally, machine learning =-=[8,9]-=-). However, the framework that is ultimately deployed may not exactly be a MapReduce implementation, mainly due to the difficulties surrounding centralized MapReduce masters in the airborne environmen...

Citation Context

...rvice, especially since no physical human intervention is possible during flight. Timely status information is therefore crucial for better decision-making by automated data center management systems =-=[85]-=- and for maintaining what amounts to a geographically-distributed, container-based data center [86].Chapter 5: Low-Overhead Status Monitoring using Compressive Sensing 131 Second, as UAVs are recruit...

Citation Context

...s just a function of the depth of the network tree, making CS an attractive solution to cloud status monitoring. Indeed, CS has already been suggested as a tool for data center temperature monitoring =-=[93]-=-. In this chapter, we propose a general architecture that allows this idea to be easily extended to other classes of applications including, e.g., CPU load monitoring for VM load spreading, per-flow b...

Citation Context

...user’s error tolerance). The common thread underlying these efforts is that they attempt to jointly optimize data correlation and routing. Unfortunately, this is a difficult problem: Cristescu et al. =-=[107]-=- show that this joint optimization of rate allocation and tree building is NP-hard. CloudSense differs in several ways from these efforts. First, the above works often rely on or assume a spatial corr...

Citation Context

...Us) as well as its wide use in a variety of application domains, ranging from indexing for web search [5] to scientific data analysis for high-energy particle physics [22] and computational chemistry =-=[23]-=- to machine learning in recommendation systems [24] and image analysis [7]. Furthermore, the traditional MapReduce deployment onto large commodity clusters has made big data processing economical for ...

Citation Context

...adio wave propagation such as path loss, fading, antenna patterns and the capture effect. We base our discussion of path loss and fading on the treatment of the subjects by Rappaport [13] and Parsons =-=[14]-=-, and of antenna patterns on that by Balanis [15]. The interested reader is referred to these texts for further details. 2.1.1 Path Loss As a receiver moves away from a transmitter, the power of the s...

Citation Context

...ibe in detail how random linear network codes work in practice. In particular, we describe the implementation of random linear network coding we will use in Chapter 4, which is similar to other works =-=[35, 36]-=-. We assume an arbitrary multihop network in which one node is a source and another a destination. The source generates an infinite stream of native packets destined for the destination. We assume nat...

Citation Context

... In both these systems, nodes promiscuously listen for coded packets from neighbors. Naturally, efforts in file sharing and content distribution led to work on multicasting over wireless. In CodeCast =-=[80]-=-, Park et al. consider the controlled loss, bounded delay multicast problem, which is applicable to multicasting data with deadlines, e.g., video. Loss is corrected by piggybacking the generation rank...

Citation Context

...es that perform routing by requiring a minimum distance from neighbors. Network coding has also been studied in the context of transport layer protocols such as TCP. Most notably, Sundararajan et al. =-=[81]-=- have proposed a network coding modification to TCP (called TCP/NC) that recasts TCP ACKs to acknowledge rank increase rather than sequence numbers, since standard TCP does not allow the destination t...

Citation Context

...erence in RSS is small. This phenomenon is known as the capture effect and has been known to occur in a wide variety of transceivers besides 802.11, such as ALOHA network [16], FM [17], and Bluetooth =-=[18]-=-. The difference in RSS between two competing transmitter’s frames can be viewed as a signal-to-noise ratio (SNR), with the stronger transmission considered the signal and the weaker one the noise1 . ...

Citation Context

...witches as well as in hardware. As follow-on work, we will be looking into implementing CloudSense switches using the NetF-Chapter 5: Low-Overhead Status Monitoring using Compressive Sensing 150 PGA =-=[98]-=- platform. 5.5 Applicability to the Airborne Setting The advantages that compressive sensing brings to the airborne compute cloud are quite promising. Below, we offer some informal arguments on how Cl...

Citation Context

... aside from these practical concerns, one technical challenge to applying MU-MIMO in the UAV setting is symbol-level synchronization (on the order of microseconds) of geographically distributed users =-=[70, 71]-=-. This problem arises because frame transmission must be aligned across users in order for each to be decoded successfully at the receiver. Existing WiMax and LTE systems have a ready solution in whic...

Citation Context

...aracterizations of the ground-to-UAV wireless channel. In the literature, there have been few efforts to characterize the performance of UAV wireless ad-hoc networks in the field. Those that do exist =-=[11, 12]-=- have focused on characterizing the point-to-point performance of air-to-air and air-to-ground wireless channels. In part, this is because these previous studies were mainly concerned with in-air rout...

Citation Context

... own SNR thresholds for capture. For 802.11 radios, the capture threshold can be <1dB, but is typically a function of the PHY bit rate, with higher bit rates requiring a higher threshold. Recent work =-=[19,20]-=- in 802.11 networks have shown that capture occurs under three distinct scenarios and that some of these cases have different SNR thresholds. For simplicity, we describe these scenarios using only two...

Citation Context

...us, low-overhead, in-network compression of status reports via compressive sensing. Our prototype design dovetails with recent work on programmable switches, such as the SideCar [90] and ServerSwitch =-=[91]-=- projects. In fact, for our prototype, we assume a hardware setup similar to SideCar: a CloudSense switch is comprised of a commodity switch connected to a general-purpose sidecar processor on a speci...

Citation Context

...ctions required some understanding of data fields or types on which they operated. Also, the compression achieved by these aggregation methods was, by definition, lossy. In contrast, the AIDA project =-=[103]-=- attempted to provide a lossless compression mechanism by concatenating multiple packets heldChapter 5: Low-Overhead Status Monitoring using Compressive Sensing 160 in a sensor’s output queue, subjec...

Citation Context

...nformation. Whenever a node notices neighbors missing chunks of a file it has, it codes all of its chunks together and broadcasts the coded chunk to help fill gaps at all of its neighbors. Lee et al. =-=[78]-=- formalized this idea into a BitTorrent-like [79] protocol called CodeTorrent for file sharing in simulated vehicular networks. In both these systems, nodes promiscuously listen for coded packets from...

Citation Context

...ies could be adapted to the UAV scenario, and (3) it was designed to process large volumes of data effectively and has been used effectively in application domains of interest (e.g., image clustering =-=[7]-=- and, more generally, machine learning [8,9]). However, the framework that is ultimately deployed may not exactly be a MapReduce implementation, mainly due to the difficulties surrounding centralized ...

Citation Context

... Bob, who can decrypt it via ℓ1-minimization after successful key exchange. But, suppose an eavesdropper Eve intercepts the ciphertext y. Is the message Alice sent Bob still secure? Rachlin and Baron =-=[99]-=- have proved that compressive sensing does not satisfy Shannon’s notion of perfect secrecy but nonetheless can be considered computationally secure since guessing Φ correctly is NP-hard. Interestingly...

Citation Context

...S), even if the difference in RSS is small. This phenomenon is known as the capture effect and has been known to occur in a wide variety of transceivers besides 802.11, such as ALOHA network [16], FM =-=[17]-=-, and Bluetooth [18]. The difference in RSS between two competing transmitter’s frames can be viewed as a signal-to-noise ratio (SNR), with the stronger transmission considered the signal and the weak...

Citation Context

...since coded packets are innovative for all recipients with high probability. Others have looked at using network coding to improve peer-to-peer file sharing in wireless mesh networks. Al Hamra et al. =-=[77]-=- proposed a selective forwarding protocolChapter 4: Improved Data Transport Using Network Coding 121 based on neighbor content information. Whenever a node notices neighbors missing chunks of a file ...

Citation Context

...aracterizations of the ground-to-UAV wireless channel. In the literature, there have been few efforts to characterize the performance of UAV wireless ad-hoc networks in the field. Those that do exist =-=[11, 12]-=- have focused on characterizing the point-to-point performance of air-to-air and air-to-ground wireless channels. In part, this is because these previous studies were mainly concerned with in-air rout...

Citation Context

...reshold of packet loss is reached. Then, the next best AP/beam pattern pair is selected and so on until none are left, at which point a new probing phase starts. In follow-on work, Subramanian et al. =-=[65]-=- extend MobiSteer to the vehicle-to-vehicle (V2V) scenario, where the vehicles exchange GPS coordinates and use the beam pattern that corresponds best to LOS 6 “Spatial multiplexing” is defined as the...

Citation Context

...on process. Below, we will discuss the basic formulation of the high-rate sensing problem and the mechanics of compressive sensing encoding and decoding, as they are typically found in the literature =-=[41, 42]-=-. Later, in Chapter 5 we will explore how CS is a natural fit for the distributed source setting. 2.4.1 The High-Rate Sensing Problem Consider a real-valued, finite-length, discrete-time signal repres...

Citation Context

...ifference between the vehicular beam steering problem in MobiSteer [64] and its UAV analog: the location and movement of an UAV is much less predictable than that of a car. Subsequently, Palat et al. =-=[74]-=- proposed cooperatively using single antenna elements in an UAV swarm as a virtual MIMO antenna array to improve data relay throughput between disconnected ground nodes via spatial diversity and multi...

Citation Context

...V extension focus on steering a single beam towards a single target and do not consider shaping the beam in multiple directions towards different APs for increased spatial diversity. The R2D2 project =-=[66]-=- investigated this shortcoming by considering the trade-off between directionality (narrowness of beams) and diversity (simultaneous coverage of multiple APs). Instead of using a single narrow beam, R...

Citation Context

... that enables continuous, low-overhead, in-network compression of status reports via compressive sensing. Our prototype design dovetails with recent work on programmable switches, such as the SideCar =-=[90]-=- and ServerSwitch [91] projects. In fact, for our prototype, we assume a hardware setup similar to SideCar: a CloudSense switch is comprised of a commodity switch connected to a general-purpose sideca...

Citation Context

...t of time. This behavior is also observed in MIMO radios—the maximum throughput for a 3x3 MIMO configuration is often advertised to be 300Mbps, but average achievable throughput is between 80–100Mbps =-=[60,61]-=-—yet, MIMO gains have been sufficiently large to vault the technology into commodity status. Second, inhospitable UAV deployment scenarios often thrust particularly harsh requirements onto the applica...

Citation Context

...d from multiple sources in the network and sent up a collection tree. En route, RNC is used to compress these data as well as realize the throughput gains from network coding. Recently, Nguyen et al. =-=[113]-=- have even implemented NetCompress, a Mote-based system similar to RNC that overcomes some practical issues such as header overhead. What RNC and NetCompress propose is essentially inter-flow compress...

Citation Context

...n airborne compute cloud enables flexible, dynamic deployment of ad-hoc network infrastructure because the aircraft can fly to any location with relative ease. For example, UAVs can act as data mules =-=[6]-=-, thereby reducing the need for labor-intensive (and possibly dangerous) deployments of terrestrial, fixed network infrastructures for backhauling data from field sensors. Second, the airborne compute...

Citation Context

...antage of any chance correlations across clusters. Other efforts have made the same observation of spatial correlation of sensor data and have used clustering techniques in similar ways. Gupta et al. =-=[106]-=- proposed to reduce traffic by selecting a subset of “representative” sensors (i.e., cluster heads) to transmit, subject to satisfying a route-to-sink connectivity constraint. Yoon et al. proposed CAG...

Citation Context

...of decoding accuracy, of course). The difference between the two methods stems from the fact that network coding encodes data over finite fields, whereas compressive sensing does so over a real field =-=[112]-=-. Katti et al. [112] have proposed Real Network Codes (RNC), in which all the finite field operations found in traditional algebraic network codes are essentially replaced with real field operations. ...

Citation Context

...t of time. This behavior is also observed in MIMO radios—the maximum throughput for a 3x3 MIMO configuration is often advertised to be 300Mbps, but average achievable throughput is between 80–100Mbps =-=[60,61]-=-—yet, MIMO gains have been sufficiently large to vault the technology into commodity status. Second, inhospitable UAV deployment scenarios often thrust particularly harsh requirements onto the applica...

Citation Context

...eceiver. Existing WiMax and LTE systems have a ready solution in which they rely on the base station to impose slotted transmission scheduling on users through the use of a reference signal, or pilot =-=[72]-=-. This mechanism makes sense for WiMax and LTE since the base station is fixed and reliable, and might even work in limited UAV settings, e.g., if a base station were mounted on a single UAV. However,...

Citation Context

...ed on hardware that has already been deployed in the field. Smart Antennas for UAVs. There is emerging interest in employing smart antenna technology and principles in the UAV setting. Breheny et al. =-=[73]-=- were among the first to propose using a swarm of UAVs as a virtual antenna array to improve swarm-tosatellite signal strength through cooperative beam steering on the UAVs. The authorsChapter 4: Imp...

Citation Context

...range in a two-hop UAV swarm relay system even under turbulence (i.e., position error) and Doppler shift assumptions, as long as low carrier frequencies were used (VHF and below). More recent efforts =-=[75,76]-=- have focused on using Kalman filters to predict the future locations of cooperating mobile ground transmitters and optimizing an UAV’s flight path to maximize aggregate uplink throughput. With FlowCo...

Citation Context

...ing data an upper bound for the throughput of a transmitter selection strategy that probes and recalculates the best transmitter once every 100 packets (approximately 1Hz, based on prior work by Vlah =-=[59]-=-). For Regions A and B, TCP+LRR average throughput deviates only 1% and 2% from the upper bounds, respectively. This result is to be expected because our data show (Table 4.1) that there is always a d...

Citation Context

...range in a two-hop UAV swarm relay system even under turbulence (i.e., position error) and Doppler shift assumptions, as long as low carrier frequencies were used (VHF and below). More recent efforts =-=[75,76]-=- have focused on using Kalman filters to predict the future locations of cooperating mobile ground transmitters and optimizing an UAV’s flight path to maximize aggregate uplink throughput. With FlowCo...

Citation Context

...ed in capability and can now output gigabytes of data per hour, meaning that the sensor data volume from military applications alone is projected to be on the order of yottabytes (1024 bytes) by 2015 =-=[2]-=-. This rapid growth in the volume of sensor data has been aptly termed the “data deluge” in the literature and has naturally become a significant concern, especially as it has far outstripped the rate...

Citation Context

...28.8% 4 52.5% 18.3% 29.2% Table 3.2: Percent of flight traces with full loss, intermediate signal, and no loss, determined using a 20-packet window. tocols such as Transmission Control Protocol (TCP) =-=[49, 50]-=- which are sensitive to packet loss patterns. Secondly, the significant presence of intermediate loss regions (16.1–18.6% in Table 3.2) indicates that we could obtain substantial gains through the use...

Citation Context

...the sheer volume of sensor data generated globally. In 2006, the total data rate of phenomenology sensors (e.g., electro-optical, infrared, radar) by the U.S. military amounted to just 270GB per year =-=[1]-=-. In the time since, sensors have increased in capability and can now output gigabytes of data per hour, meaning that the sensor data volume from military applications alone is projected to be on the ...

Citation Context

... CloudSense TOR switches always buffer the latest report and 1 ΦA and ΦB could be the same if we choose a proper sparsifying transform that distinguishes 〈x1, x2, x3〉 from 〈x4, x5, x6〉 to some degree =-=[92]-=-. In our setting, since the sparsifying transform is assumed to be the identity, we must use distinct Φj. However, it can be desirable to use the same Φj since this will reduce the total number of var...