Due to the advances in deep learning, visually-aware recommender systems (RS) have recently attracted increased research interest. Such systems combine collaborative signals with images, usually represented as feature vectors outputted by pre-trained image models. Since item catalogs can be huge, recommendation service providers often rely on images that are supplied by the item providers. In this work, we show that relying on such external sources can make an RS vulnerable to attacks, where the goal of the attacker is to unfairly promote certain pushed items. Specifically, we demonstrate how a new visual attack model can effectively influence the item scores and rankings in a black-box approach, i.e., without knowing the parameters of the model. The main underlying idea is to systematically create small human-imperceptible perturbations of the pushed item image and to devise appropriate gradient approximation methods to incrementally raise the pushed item's score. Experimental evaluations on two datasets show that the novel attack model is effective even when the contribution of the visual features to the overall performance of the recommender system is modest.

In this paper, we introduce a new channel model termed as the q-ary multi-bit channel. This channel models a memory device, where q-ary symbols (q = 2s) are stored in the form of current/voltage levels. The symbols are read in a measurement process, which provides a symbol bit in each measurement step, starting from the most significant bit. An error event occurs when not all the symbol bits are known. To deal with such error events, we use GF(q) lowdensity parity-check (LDPC) codes and analyze their decoding performance. We start with iterative-decoding threshold analysis and derive optimal edge-label distributions for maximizing the decoding threshold. We later move to a finite-length iterative decoding analysis and propose an edge-labeling algorithm for the improved decoding performance. We then provide a finite-length maximum-likelihood decoding analysis for both the standard non-binary random ensemble and LDPC ensembles. Finally, we demonstrate by simulations that the proposed edge-labeling algorithm improves the finite-length decoding performance by orders of magnitude.

Recent advances in coding for distributed storage systems have reignited the interest in scalar codes over extension fields. In parallel, the rise of large-scale distributed systems has motivated the study of computing in the presence of stragglers, i.e., servers that are slow to respond or unavailable. This paper addresses storage systems that employ linear codes over extension fields. A common task in such systems is the reconstruction of the entire dataset using sequential symbol transmissions from multiple servers, which are received concurrently at a central data collector. However, a key bottleneck in the reconstruction process is the possible presence of stragglers, which may result in excessive latency. To mitigate the straggler effect, the reconstruction should be possible given any sufficiently large set of sequentially received symbols, regardless of their source. In what follows, an algebraic framework for this scenario is given, and a number of explicit constructions are provided. Our main result is a construction that uses a recursive composition of generalized Reed-Solomon codes over smaller fields. In addition, we show links of this problem to Gabidulin codes and to universally decodable matrices.

In this paper, we address the finite-length decoding performance of LDPC codes over the q-ary multi-bit channel (QMBC). The QMBC is defined over the full q-ary symbols, while addressing the differences in reliability between the bits composing the symbols. We show that unlike the binary erasure channel, the QMBC iterative decoder does not necessarily halt at stopping sets. Instead, its performance depends on the edge-label configuration of the LDPC code graph. We characterize good edge-label configurations, and propose an edge-labeling algorithm for improved iterative-decoding performance. We then provide finite-length maximum-likelihood decoding analysis for both the standard non-binary random ensemble and LDPC ensembles. Finally, simulations are presented to demonstrate the advantages of the proposed edge-labeling algorithm.

Network switches and routers need to serve packet writes and reads at rates that challenge the most advanced memory technologies. As a result, scaling the switching rates is commonly done by parallelizing the packet I/Os using multiple memory units. For improved read rates, packets can be coded upon write, thus giving more flexibility at read time to achieve higher utilization of the memory units. This paper presents a detailed study of coded network switches, and in particular, how to design them to maximize the throughput advantages over standard uncoded switches. Toward that objective, the paper contributes a variety of algorithmic and analytical tools to improve and evaluate the throughput performance. The most interesting finding of this paper is that the placement of packets in the switch memory is the key to both high performance and algorithmic efficiency. One particular placement policy we call “design placement” is shown to enjoy the best combination of throughput performance and implementation feasibility.

Coded switches write incoming packets with redundancy to increase the flexibility to read them later without contention. An important question pertaining to coded switches is what policy to follow when placing the coded packets in the switch memory. We study this question by proposing two such placement policies: cyclic placement and (block-) design placement. We show that these policies offer many advantages in switching throughput, algorithmic efficiency, and analysis amenability.

With the increasing demand for network bandwidth, network switches face the challenge of serving growing data rates. To parallelize the process of writing and reading packets to the switch memory, multiple memory units (MUs) are deployed in parallel in the switch fabric. However, memory contention may occur if packets requested to read happen to share one or more MUs, due to memory bandwidth limitations. Avoiding such contention in the write stage is limited as the reading schedule of packets is not known upon arrival of the packets to the switch. Thus, efficient packet placement and read policies are required.

In this paper, we develop a new channel model, which we name the q-ary partial erasure channel (QPEC). The QPEC has a q-ary input, and its output is either the input symbol or a set of M (2 ≤ M ≤ q) symbols, containing the input symbol. This channel serves as a generalization to the binary erasure channel and mimics situations when a symbol output from the channel is known only partially; that is, the output symbol contains some ambiguity, but is not fully erased. This type of channel is motivated by non-volatile memory multi-level read channels. In such channels, the readout is obtained by a sequence of current/voltage measurements, which may terminate with a partial knowledge of the stored level. Our investigation is concentrated on the performance of low-density parity-check (LDPC) codes when used over this channel, thanks to their low decoding complexity using belief propagation. We provide the exact QPEC density-evolution equations that govern the decoding process, and suggest a cardinality-based approximation as a proxy. We then provide several bounds and approximations on the proxy density evolutions, and verify their tightness through numerical experiments. Finally, we provide tools for the practical design of LDPC codes for use over the QPEC.

Today, designers of network processors strive to keep the packet reception and transmission orders identical, and therefore avoid any possible out-of-order transmission. However, the development of new features in advanced network processors has resulted in increasingly parallel architectures and increasingly heterogeneous packet processing times, leading to large reordering delays. In this paper, we introduce novel scalable scheduling algorithms for preserving flow order in parallel multi-core network processors. We show how these algorithms can reduce reordering delay while adapting to any load-balancing algorithm and keeping a low implementation complexity overhead. To do so, we use the observation that all packets in a given flow have similar processing requirements and can be described with a constant number of logical processing phases. We further define three possible knowledge frameworks of the time when a network processor learns about these logical phases, and deduce appropriate algorithms for each of these frameworks. Finally, we model our proposed algorithms and simulate them under both synthetic traffic and real-life traces, and show that they significantly outperform past approaches.

Network services, such as security, load-balancing, and monitoring, are an indisputable part of modern networking infrastructure and are traditionally realized as specialized appliances or middleboxes. Middleboxes complicate the management, the deployment, and the operations of the entire network. Moreover, they induce network performance issues and scalability limitations by requiring huge amounts of traffic to be, often sub-optimally redirected, and sometimes redundantly processed. Recent trends of server virtualization and Network Function Vir-tualization (NFV) exacerbate these scalability and performance issues. In this paper, we present EnforSDN — a new management approach that exploits SDN principles to decouple the policy resolution layer from the policy enforcement layer in network service appliances. Our approach improves the enforcement management, network utilization and communication latency, without compromising the policy and the functionality of the network. Using emulated SDN-based data center environment, we demonstrate higher throughput and lower latency achieved with EnforSDN, as compared to a baseline SDN network. In addition, we show that EnforSDN reduces the overall network appliances load, as well as the forwarding tables size.

Network Function Virtualization (NFV) is a new networking paradigm where network functions are executed on commodity servers located in small cloud nodes distributed across the network, and where software defined mechanisms are used to control the network flows. This paradigm is a major turning point in the evolution of networking, as it introduces high expectations for enhanced economical network services, as well as major technical challenges. In this paper, we address one of the main technical challenges in this domain: the actual placement of the virtual functions within the physical network. This placement has a critical impact on the performance of the network, as well as on its reliability and operation cost. We perform a thorough study of the NFV location problem, show that it introduces a new type of optimization problems, and provide near optimal approximation algorithms guaranteeing a placement with theoretically proven performance. The performance of the solution is evaluated with respect to two measures: the distance cost between the clients and the virtual functions by which they are served, as well as the setup costs of these functions. We provide bi-criteria solutions reaching constant approximation factors with respect to the overall performance, and adhering to the capacity constraints of the networking infrastructure by a constant factor as well. Finally, using extensive simulations, we show that the proposed algorithms perform well in many realistic scenarios.

Network switches and routers need to serve packet writes and reads at rates that challenge the most advanced memory technologies. As a result, scaling the switching rates is commonly done by parallelizing the packet I/Os using multiple memory units. For improved read rates, packets can be coded with an [n,k] MDS code, thus giving more flexibility at read time to achieve higher utilization of the memory units. In the paper, we study the usage of [n,k] MDS codes in a switching environment. In particular, we study the algorithmic problem of maximizing the instantaneous read rate given a set of packet requests and the current layout of the coded packets in memory. The most interesting results from practical standpoint show how the complexity of reaching optimal read rate depends strongly on the writing policy of the coded packets.

In this paper, we discuss practical design of low-density parity-check (LDPC) codes for the q-ary partial erasure channel (QPEC). This channel is an extension of the binary erasure channel (BEC), where partial information on the output is available. We provide a linear programming (LP) optimization for the design of good degree distributions, and compare our code design results to codes obtained using an LP optimization formulated for the BEC. We show superior performance in terms of code rate and complexity, when designing an LDPC code for a desired decoding threshold.

Today, designers of network processors strive to keep the packet reception and transmission orders identical, and therefore avoid any possible out-of-order transmission. However, the development of new features in advanced network processors has resulted in increasingly parallel architectures and increasingly heterogeneous packet processing times, leading to large reordering delays. In this paper, we introduce novel scalable scheduling algorithms for preserving flow order in parallel multi-core network processors. We show how these algorithms can reduce reordering delay while adapting to any load-balancing algorithm and keeping a low implementation complexity overhead. To do so, we use the observation that all packets in a given flow have similar processing requirements and can be described with a constant number of logical processing phases. We further define three possible knowledge frameworks of the time when a network processor learns about these logical phases, and deduce appropriate algorithms for each of these frameworks.

Software Defined Networks (SDNs) are becoming the leading technology behind many traffic engineering solutions, both for backbone and data-center networks, since it allows a central controller to globally plan the path of the flows according to the operator's objective. Nevertheless, networking devices' forwarding table is a limited and expensive resource (e.g., TCAM-based switches) which should thus be considered upon configuring the network. In this paper, we concentrate on satisfying global network objectives, such as maximum flow, in environments where the size of the forwarding table in network devices is limited. We formulate this problem as an (NP-hard) optimization problem and present approximation algorithms for it. We show through extensive simulations that practical use of our algorithms (both in Data Center and backbone scenarios) result in a significant reduction (factor 3) in forwarding table size, while having a small effect on the global objective (maximum flow).

In this paper, we develop a new channel model, which we name the q-ary partial erasure channel (QPEC). QPEC has a q-ary input, and its output is either one symbol or a set of M possible values. This channel mimics situations when current/voltage levels in measurement channels are only partially known, due to high read rates or imperfect current/voltage sensing. Our investigation is concentrated on the performance of low-density parity-check (LDPC) codes when used over this channel, due to their low decoding complexity with iterative-decoding algorithms. We give the density evolution equations of this channel, and develop its decoding-threshold analysis. Part of the analysis shows that finding the exact decoding threshold efficiently lies upon a solution to an open problem in additive combinatorics. For this part, we give bounds and approximations.

Datacenter-based Cloud computing has induced new disruptive trends in networking, key among which is network virtualization. Software-Defined Networking overlays aim to improve the efficiency of the next generation multitenant datacenters. While early overlay prototypes are already available, they focus mainly on core functionality, with little being known yet about their impact on the system level performance. Using query completion time as our primary performance metric, we evaluate the overlay network impact on two representative datacenter workloads, Partition/Aggregate and 3-Tier. We measure how much performance is traded for overlay's benefits in manageability, security and policing. Finally, we aim to assist the datacenter architects by providing a detailed evaluation of the key overlay choices, all made possible by our accurate cross-layer hybrid/mesoscale simulation platform.

Overlay routing is a very attractive scheme that allows improving certain properties of the routing (such as delay or TCP throughput) without the need to change the standards of the current underlying routing. However, deploying overlay routing requires the placement and maintenance of overlay infrastructure. This gives rise to the following optimization problem: Find a minimal set of overlay nodes such that the required routing properties are satisfied. In this paper, we rigorously study this optimization problem. We show that it is NP-hard and derive a nontrivial approximation algorithm for it, where the approximation ratio depends on specific properties of the problem at hand. We examine the practical aspects of the scheme by evaluating the gain one can get over several real scenarios. The first one is BGP routing, and we show, using up-to-date data reflecting the current BGP routing policy in the Internet, that a relative small number of less than 100 relay servers is sufficient to enable routing over shortest paths from a single source to all autonomous systems (ASs), reducing the average path length of inflated paths by 40%. We also demonstrate that the scheme is very useful for TCP performance improvement (results in an almost optimal placement of overlay nodes) and for Voice-over-IP (VoIP) applications where a small number of overlay nodes can significantly reduce the maximal peer-to-peer delay.

The recent growing popularity of cloud-based solutions and the variety of new applications present new challenges for cloud management and resource utilization. In this paper we concentrate on the networking aspect and consider the placement problem of virtual machines (VMs) of applications with intense bandwidth requirements. Optimizing the available network bandwidth is far more complex than optimizing resources like memory or CPU, since every network link may be used by many physical hosts and thus by the VMs residing in these hosts. We focus on maximizing the benefit from the overall communication sent by the VMs to a single designated point in the data center (called the root). This is the typical case when considering a storage area network of applications with intense storage requirements. We formulate a bandwidth-constrained VM placement optimization problem that models this setting. This problem is NP hard, and we present a polynomial-time constant approximation algorithm for its most general version, in which hosts are connected to the root by a general network graph. For more practical cases, in which the network topology is a tree and the revenue is a simple function of the allocated bandwidth, we present improved approximation algorithms that are more efficient in terms of running time. We evaluate the expected performance of our proposed algorithms through a simulation study over traces from a real production data center, providing strong indications to the superiority of our proposed solutions.

In recent years, hardware-based packet classification has became an essential component in many networking devices. It often relies on ternary content-addressable memories (TCAMs), which can compare in parallel the packet header against a large set of rules. Designers of TCAMs often have to deal with unpredictable sets of rules. These result in highly variable rule expansions, and can only rely on heuristic encoding algorithms with no reasonable guarantees. In this paper, given several types of rules, we provide new upper bounds on the TCAM worst case rule expansions. In particular, we prove that a W-bit range can be encoded in W TCAM entries, improving upon the previously known bound of 2W - 5. We further prove the optimality of this bound of W for prefix encoding, using new analytical tools based on independent sets and alternating paths. Next, we generalize these lower bounds to a new class of codes called hierarchical codes that includes both binary codes and Gray codes. Last, we propose a modified TCAM architecture that can use additional logic to significantly reduce the rule expansions, both in the worst case and using real-life classification databases.

Overlay routing in a very attractive scheme that allows improving certain properties of the routing without the need to change the standards of the current underlying routing. However, deploying overlay routing requires the placement and maintenance of overlay infrastructure. This gives rise to the following optimization problem: find a minimal set of overlay nodes such that the required routing properties are satisfied. In this paper we rigorously study this optimization problem. We show that it is NP hard and derive a non-trivial approximation algorithm for it, where the approximation ratio depends on specific properties of the problem at hand. We examine the practical aspects of the scheme by evaluating the gain one can get over two real scenarios. The first one is BGP routing, and we show, using up-to-date data reflecting the current BGP routing policy in the Internet, that a relative small number of less than 100 relay servers are sufficient to enable routing over shorter paths from a single source to all ASes, reducing the average path length of inflated paths by 40%. We also demonstrate that using the scheme for TCP performance improvement, results in an almost optimal placement of overlay nodes.

In recent years, hardware based packet classification has became an essential component in many networking devices. Ternary Content-Addressable Memories (TCAMs) are one of the most popular solutions in this domain, allowing to compare in parallel the packet header against a large set of rules, and to retrieve the first match. However, using TCAM to match a range of values is much more problematic and dramatically reduces the cost effectiveness of the solution. In this paper we study ways to use simple built-in TCAM mechanisms in order to increase the efficiency of range coverage. While current techniques have a worst expansion ratio of 2W-4, we present an efficient algorithm enabling to encode any range with at most W TCAM entries (where W in the number of bits), without using additional processing, extra bits, and without any external encoding. The same paradigm can be applied to multiple raging rules as well, resulting in significant improvement over current known techniques. Moreover, our simulation results indicate that these techniques can be used to reduce the actual TCAM size of hardware networking devices under realistic scenarios.

Monitoring is an essential part of any self-adaptive management loop. While providing the necessary information for making management decisions, monitoring itself incurs a cost in terms of the system and network resources committed to this management task. Thus, one can pose a generic question: what is the right amount of monitoring that maximizes its utility for management? This question turns out to be difficult to answer in general. In this paper we focus on quantifying the utility of monitoring for self-adaptive load sharing, where a stream of jobs arrives at a collection of n identical servers. We propose a novel model, that we dubbed an Extended Supermarket Model (ESM) to study the tradeoff between the usefulness of the monitoring information and the cost of obtaining it. We show that for each service request rate, there exists an optimal number of servers that should be monitored to obtain minimal average service time at an optimal cost. Using these findings, we present self-adaptive load-sharing algorithms both for centralized and fully distributed settings and evaluate these algorithms using simulations and a real testbed. Our results show that in realistic scenarios, where monitoring cost is not negligible, the self-adaptive load balancing is clearly superior to any cost-oblivious load-sharing mechanisms. We also demonstrate that in a fully distributed setting, where no dedicated monitoring component is employed, our self-adaptive heuristics perform very well with respect to the current common practice.

Many network applications that need to distribute content and data to a large number of clients use a hybrid scheme in which one (or more) multicast channel is used in parallel to a unicast dissemination. This way the application can distribute data using one of its available multicast channels or by sending one or more unicast transmissions. In such a model the utilization of the multicast channels is critical for the overall performance of the system. We study the scheduling algorithm of the sender in such a model. We describe this scheduling problem as an optimization problem where the objective is to maximize the utilization of the multicast channel. Our model captures the fact that it may be beneficial to multicast an object more than once (e.g., page update). Thus, the benefit depends, among other things, on the last time the object was sent, which makes the problem much more complex than previous related scheduling problems. We show that our problem is NP-hard. Then, using the local ratio technique we obtain a 4-approximation algorithm for the case where the objects are of fixed size and a 10-approximation algorithm for the general case. We also consider a special case which may be of practical interest, and prove that a simple greedy algorithm is a 3-approximation algorithm in this case.

The Internet connectivity in the autonomous system (AS) level reflects the commercial relationship between ASes. A connection between two ASes could be of typecustomer-providerwhen one AS is a provider of the other AS, or of typepeer-peer, if they are peering ASes. This commercial relationship induces a global hierarchical structure which is a key ingredient in the ability to understand the topological structure of the AS connectivity graph. Unfortunately, it is very difficult to collect data regarding the actual type of the relationships between ASes, and in general this information is not part of the collected AS connectivity data. The Type of Relationship (ToR) problem attempts to address this shortcoming, by inferring the type of relationship between connected ASes based on their routing policies. However, the approaches presented so far are local in nature and do not capture the global hierarchical structure. In this work we define a novel way to infer this type of relationship from the collected data, taking into consideration both local policies and global hierarchy constrains. We define the Acyclic Type of RelationshipAToRproblem that captures this global hierarchy and present an efficient algorithm that allows determining if there is a hierarchical assignment without invalid paths. We then show that the related general optimization problem is NP-complete and present a 2/3 approximation algorithm where the objective function is to minimize the total number of local policy mismatches. We support our approach by extensive experiments and simulation results showing that our algorithms classify the type of relationship between ASes much better than all previous algorithms.

Abstract — The topological structure of the Internet infrastructure is an important and interesting subject that attracted significant research attention in the last few years. Apart from the pure intellectual challenge of understanding a very big, complex, and ever evolving system, knowing the structure of the Internet topology is very important for developing and studying new protocols and algorithms. Starting with the fundamental work of Falostous et. al, a considerable amount of work was done recently in this field, improving our knowledge and understanding of the Internet structure. However, one basic problem is still unanswered: how big is the Internet. In the AS level this means: how many peering relations exist between ASs. Finding this number is hard since there is no direct way to retrieve information from all nodes regarding their direct neighbors, and all our knowledge is based on sampling processes. Thus, it is very difficult to characterize the Internet since it may well be the case that this characterization is a result of the sampling process, and it does not hold for the “real ” Internet. In this paper we attack this problem by suggesting a novel usage of the measurements themselves in order to infer information regarding the whole system. In other words, in addition to looking at the overall graph that is generated from the union of the data obtained by performing many measurements, we consider the actual different measurements and the amount of new data obtained in each of them with respect to the previous collected data. Using the second moment allows us to reach conclusions regarding the structure of the system we are measuring, and in particular to estimate its total size. We present strong evidence to the fact that a considerable amount (at least 35%) of the links in the AS level are still to be unveiled. Our findings indicate that almost all these missing links are of type peer-peer, and we provide novel insight regarding the structure of the AS connectivity map with respect to the peering type. I.

The rapid growth of wireless and cellular networks, and the high availability of small communication devices, such as PDAs, brings us faster than ever to the point where context aware services (CASs) are becoming a commodity. In order to allow fast and efficient development, deployment, and management of such services, a global system that allows the services to gain access to the context information needs to be created, maintained, and managed. We study the requirements for such a context distribution system. We deal with the architectural decisions regarding the definition of context items and the way context information becomes available to the CASs, and also the algorithmic aspects of disseminating this information. We demonstrate the advantages of the architecture and the proposed information dissemination algorithms by conducting a simulation study under realistic practical assumptions. Our results indicate that a modular approach in which context information is provided in many network locations by brokers through an open simple API is both powerful enough to provide the needed context information, and simple enough to be easily implemented.