Fluent Essays

read paper and provide feedback/review in the format attached. Attaching sample reference. NOTE: Please put X to show your selection Type of this paper Research Survey Tutorial Speculative Others Your Choice X Evaluation: Very Poor Poor Average Good Very Good Significance of the main idea(s) X Originality X Technical quality of the paper X Awareness of related work X Clarity of presentation X Organization of the manuscript X References X Paper Length X How comfortable are you in reviewing this pape r? Very Confident Confident Adequate Not Confident Not my Area Your Choice X Overall comments and changes that MUST be made before publication: · Author discusses the role of artificial intelligence and machine learning in the finance sector. · Introduction section should be improved. · Separate section for related work will be an added advantage for this paper. The authors need to inject the section of Related Work preferably after Introduction and investigate majority most of the previous work. · Nothing is novel since it is a survey article. But it can be good reference work for new researchers. · However, in order to present both the results and the conclusions as clearly as possible, we suggest to the authors of the paper the presentation of personal scientific contributions to the literature, as well as the presentation from an applicative point of view. · More references should be cited in reference section as well as body of the article. · Increase the length of the paper by making each section more comprehensive. · his work resembles an awareness of the use of artificial intelligence and its benefits in the financial field. Indeed the author in his awareness: - presents artificial intelligence in general - the financial areas in which it could be used - presents the importance of the diversity of data handled in finance. This could increase the level of risk associated with the processing of this data. The author tells us that artificial intelligence could make it possible to put in place triggers allowing to know the presence of possible risks. In these writings does not present how (in terms of algorithm - of process implemented) artificial intelligence comes to have a very important role in finance. It could be based on a practical case to support his remarks, especially since the manuscript is a succession of affirmations without associated scientific references Suggestions which would improve the quality of the paper but are NOT essential for publication:  1 - expand your bibliography 2 - carry out a good bibliographic review 3 - identify recent work carried out in connection with the role and associated results 4 - carry out a comparative study with the existing ones in technical terms of artificial intelligence used in the field of finance 5 - make scientific proposals with regard to the limits of what already exists by supporting these proposals with at least one practical case. Overall Recommendation: Strongly Reject Reject Marginally Accept Accept Strongly Accept Recommendation X PAPER REVIEW Chapter Title: A Comprehensive Survey of Energy-Efficiency Approaches in Wired Networks Authors: NOTE: Please put X to show your selection Type of this paper   Research Survey Tutorial Speculative Others Your Choice Evaluation:   Very Poor Poor Average Good Very Good Significance of the main idea(s)  Originality Technical quality of the paper  Awareness of related work Clarity of presentation  Organization of the manuscript References Paper Length Overall comments and changes that MUST be made before publication: · Please write the detailed comments on paper and review here (say in 5 – 10 bullet points) Suggestions which would improve the quality of the paper but are NOT essential for publication: · Please write the suggestions to improve the paper here (say in 2 or 3 bullet points) Overall Recommendation:   Strongly Reject Reject Marginally Accept Accept Strongly Accept Recommendation Referee’s Name and Date A COMPREHENSIVE SURVEY OF ENERGY-EFFICIENCY APPROACHES IN WIRED NETWORKS ABSTRACT Energy consumption by the network infrastructure is growing expeditiously with the rise of the Internet. Critical research efforts have been pursued by academia, industry and governments to make networks, such as the Internet, operate more energy efficiently and reduce their power consumption. This work presents an in-depth survey of the approaches to reduce energy consumption in wired networks by first categorizing existing research into broad categories and then presenting the specific techniques, research challenges, and important conclusions. At a broad level, we present five categories of approaches for energy efficiency in wired networks – (i) sleeping of network elements, (ii) link rate adaptation, (iii) proxying, (iv) store and forward, and (v) network traffic aggregation. Additionally, this survey reviews work in energy modeling and measurement, energy-related standards and metrics, and enumerates discussion points for future work and motivations. KEYWORDS Energy efficiency, energy proportionality, energy-aware protocols, wired networks 1. INTRODUCTION The Internet has proven to be one of the most important technological innovations. It acts as the primary catalyst in the global digital revolution, and is considered a public utility, along with running water and electricity. Fig. 1 compares access to running water to the Internet in the USA. Although this is not a like-for-like comparison, it is still important to note the pace of Internet proliferation when compared to other public utilities. Recent statistics estimate the global Internet adoption rate to be around 65.6\%, or 5.1 billion users [1], and these numbers are increasing rapidly. In addition to the escalation of Internet adoption, the digitization of services including over the top (OTT) video streaming, ecommerce, voice over IP (VoIP), and the Internet of things (IoT), has established pressure on service and network providers to expand their network hardware infrastructure. This network hardware expansion results in an upward trend of the energy consumed by the Internet globally. While Koomey’s law states that post-2000 the energy efficiency of computing hardware has doubled every 2.6 years, it is still slower than the increase in data traffic, which follows Moore’s law, doubling every 18 months [2]. Although the energy consumption by the wired-networking infrastructure is a small fraction of the total consumption by the information and communications technologies, the absolute numbers indicate that efforts to reduce energy consumption in computer networks are warranted [3]. Figure 1. Access to running water vs the Internet in the USA In the past two decades, there has been a surge of research dedicated to make the network infrastructure more energy efficient. The three major areas of focus have been, (i) system-oriented energy management, (ii) energy-aware network design, and (iii) energy-aware protocol design. System-oriented approaches strive to make the underlying hardware more energy-efficient using techniques like frequency scaling, dynamic voltage scaling (DVS), providing additional sleep states (like C-states in Intel) and performance states (like P-states in Intel). Energy-aware network design approaches seek to find the optimal network topologies that are most suited for energy savings while maintaining performance and reliability standards. Energy-aware protocol design approaches attempt to rethink and remodel existing protocols, and propose new protocols, to incorporate energy- awareness with the goal to provide opportunities to reduce network energy consumption. The system- oriented approaches are comparatively more mature than the other two approaches. This survey paper, therefore, excludes system-oriented approaches and focuses on network design and protocol-oriented approaches for wired networks. We study the work done in the last twenty years, while analyzing and discussing the research challenges, with the aim to present the state-of-the-art and realize important conclusions that can benefit future work in this field. The remainder of the paper is organized as follows: Section II discusses the broad categories of the existing work, Section III presents a detailed survey of the specific approaches in each category, Section IV presents work that is closely related, but are not included in Section III, and Section V presents conclusions and discussion points for future research. 2. CATEGORIZATION OF EXISTING WORK Computer networks are a critical part of the Internet infrastructure acting as highways connecting users to online services. They are conventionally designed to handle peak-traffic loads with sufficient redundancy and Quality-of-Service (QoS). Owing to the best-effort nature of the Internet Protocol (IP), network architectures and applications are developed to provide reliable transmission in case of failures. While such techniques have promoted the growth of the Internet, they are not typically energy-proportional, i.e. their energy consumption is not proportional to the traffic load. The practice of over-provisioning and observations of under-utilization and non-energy-proportional behavior during the majority of times prompted researchers to find methods to reduce the energy-consumption of wired networks. Since individual network hardware devices and their components have become increasingly reliable, and network protocols have become more robust, new techniques can be leveraged to determine better energy savings. Different parts of the wired-network infrastructure – network topology, hardware devices, electrical/optical links, protocols, and network management techniques – exhibit different runtime behaviors and offer different opportunities to lower the energy consumption. We categorize the existing work into five distinct categories: sleeping of network elements, link rate adaptation, proxying, store and forward, and network traffic aggregation, and provide an overview of them in this section. Each category targets a specific area of wired-networking with its own research challenges, implementation-specific techniques, and scope for future work. Fig. 2 shows the taxonomy of existing research in different categories. The detailed survey of specific approaches in each category is presented in section III. The concept of sleeping of network elements is based on the intuition that because computer networks are provisioned for peak load and are generally under-utilized, parts of the network or individual devices can be put to sleep during off-peak hours. This category of work was one of the earliest approaches (2003) to save network energy consumption and has seen a multitude of proposals. The work involved in this approach is two-fold: identifying/predicting low-utilization periods suitable for energy savings, and reconfiguring the network to put network elements to sleep. The target network element to be put to sleep could either be the physical device(s), or individual components of one device. Although networking devices traditionally provide limited support for sleeping, approaches in this category propose techniques which either make assumptions about the underlying hardware support, or find ways to work around this limitation. Waking up a sleeping element and understanding the impact of sleeping on network protocols were some of the research challenges in this category. Link rate adaption gained popularity as a potential approach after research proved that high link rates (1/10 Gbps) consume more energy as compared to low link rates (10/100 Mbps), and that a desirable level of performance could be achieved by running links at lower speeds. The work involved in this approach is two-fold: identifying/predicting low-utilization periods wherein the links could be operated at slower rates, and reconfiguring the network to configure specific links at lower speeds. Both reactive approaches, based on matching the current utilization with link speeds, and proactive approaches, predicting current/future traffic patterns from historical utilization data, were proposed. Research challenges in this category include understanding the tradeoffs between frequent link-rate change and performance, and the impact of these changes on protocols, especially cost-based routing and switching protocols. Proxying is based on the hypothesis that the periodic network protocol traffic does not allow end- hosts to sleep effectively. Although this approach does not directly target energy savings from the network, it involves the network elements to act as proxies to enable infrastructure-level savings. The techniques proposed in this category attempted to counter the requirements necessitating networked end-hosts to be connected and responsive even while sleeping, i.e. always-on should not mean always powered on. A survey of office and commercial equipment indicated that most of the computers were not put into sleep modes due to the above requirement of always maintaining network presence [27], and this prompted researchers to consider network proxying and find ways to offload some of the end-host responsibilities to a proxy. The placement of the proxy in the network and the tradeoff between energy savings by end-hosts sleeping and the additional energy consumption by the proxy were some of the research challenges in this area of work. Figure 2. Categorization of existing work (* indicates work appears in multiple categories) The observation that many web applications exhibit bursty traffic with high inter-arrival traffic times led to the development of store and forward techniques. Although this is not a mature area and does not have many research proposals, it is still an important area of work to study with insights for future works. While the traditional practice in network engineering is to avoid bursts using congestion avoidance and QoS queues, the insight behind the techniques in this area is to shape traffic in bursts in order to allow network elements to be put in low power/sleep mode in between these bursts. The difference between sleeping of network elements and store and forward is that techniques in this area use explicit proactive mechanisms to shape and engineer traffic with the intention to create sufficient bunching that would allow efficient sleeping, while techniques in the former aim to reactively put network elements to sleep based on utilizations. Store and forward techniques could be considered as a subset of the sleeping of network elements category, but we believe there are sufficient design differences in the two areas to warrant separate categories. Developing practical ways to implement this approach, synchronizing the bursts across the network to enable efficient sleeping, and understanding the tradeoffs between performance and energy savings were some of the research challenges in this area. Network traffic aggregation techniques are based on the assumption that networks are over- provisioned and under-utilized, and that a subset of devices and links would suffice during off-peak hours. While we do not include server load aggregation techniques – clustering compute resources onto minimal physical servers to save energy – in this area, specific techniques that involve manipulating network traffic to enable server-load aggregation are included. The difference between network aggregation techniques and sleeping of network elements is that approaches in this area attempt to proactively find minimum-power subsets of the infrastructure that can provide the desired level of performance while enabling energy savings, whereas, techniques on sleeping of network elements attempt to put network elements to sleep reactively while considering the entire topology. These techniques could be included in the first category, but we believe that the insight and the way forward is different enough to justify separate categories. Research challenges in this area focus on traffic engineering problems such as NP-complete [37] and finding methods to reduce the computing time that allows for on-the-fly solutions to efficiently save energy. 3. DETAILED SURVEY In this section we present a chronological survey of specific approaches in each category. 3.1. Sleeping of network elements Gupta and Singh [3] were among the earliest researchers to examine the energy consumption of networking devices. A previous study which analyzed the energy consumption of office and telecommunications equipment in commercial buildings revealed that Internet devices consumed 6.05 TW-h of energy in the USA [4]. This realization prompted the researchers to find ways and suggest directions to save energy going forward. They proposed sleeping of devices/interfaces as a solution and explored coordinated and uncoordinated sleeping. In coordinated sleeping, the network devices collectively made decisions on which elements to put to sleep, while in uncoordinated sleeping, devices made such decisions in isolation, independent of others. Analyzing sample traces from their autonomous system (AS), they showed that sleeping was a reasonable solution, however, modifications to existing protocols and the Internet architecture would be needed to maximize the amount of energy conserved. Soteriou and Peh [6] earmarked links interconnecting routers as a major source of energy consumption and proposed a dynamic power management policy employing on/off links. The work involved deriving a power-performance connectivity graph to identify candidate on/off links, developing a deadlock-free routing algorithm based on the information from the graph, and implementing an on/off decision mechanism. Input buffer utilization of a router was used to determine if a link could be turned off. Their results indicated a 37.5\% reduction in energy consumption for an 8-ary 2-mesh topology with a moderate latency increase. Gupta et al. extended their prior work in [5] to investigate the feasibility of sleeping in LAN devices. LAN switches are targeted as they comprise the bulk of LAN devices and therefore consume the largest amount of energy. Three different sleep models were proposed – (i) Simple sleep – wherein timers are employed to wake up a sleeping device/interface and all incoming packets during sleep time are lost, (ii) Hardware assisted sleep (HAS) – wherein the incoming packet is lost but it wakes up a sleeping device/interface, and (iii) Hardware assisted buffered sleep (HABS) – wherein the incoming packet is buffered and it wakes up a sleeping device/interface. Both HAS and HABS assumed hardware support which was not available at the time. A need to modify existing protocols (like STP and TCP) to allow for efficient sleeping was reiterated. While the work in [5] was a preliminary study to evaluate the viability of sleeping as a solution to save energy, in [7] Gupta et al. presented practical approaches to achieve it, using technology available in the hardware of the time. The target area was Ethernet interfaces and their proposal leveraged smart timer-based Ethernet transceivers that automatically turned off when no power was detected on the other end. The algorithms to decide when to turn off a link included – (i) On/Off-1 – wherein the upstream interface of a link determines if the link can be put to sleep by estimating the number of packet arrivals in a time period t, and finding the maximum value of t for which the probability of packet arrivals being less than a buffer threshold is less than 10\%, and (ii) On/Off-2 – which is a modification of the previous algorithm, wherein instead of both the upstream and downstream interfaces waking up after time period t and reevaluating if they can go back to sleep, only the upstream interface wakes up and runs the algorithm, to allow for higher energy savings. The simulation results on real-world traces indicated that 37\% of the interfaces could be put to sleep anywhere from 40-98\% of the time. Another conclusion from their work was that sleeping interfaces allowed portions of the internal switching fabric to be put to sleep as well. Nedevschi et al. [8] studied opportunistic sleeping in which link interfaces sleep when no packet arrivals are observed for a period of time. This approach assumed hardware support of wake-on- arrival – the circuitry that sensed traffic on the interface and has the capability to wake up the device. The simulation results on Abilene showed that opportunistic sleeping was suitable for LANs with high idle times, however, it was not suitable for high-speed links with low idle times. The percentages of time links could sleep with either constant bit rate traffic or bursty traffic were minimal, proving that this approach was suitable only in the specific LAN scenarios. Following similar intuitions as before, Ananthanarayanan and Katz [9] proposed two energy-saving schemes for switches and studied the tradeoffs between performance and energy consumption. Their schemes fall under the umbrella of uncoordinated sleeping in the way that they are run on each switch independently. The concept of shadow port is introduced wherein a shadow port is associated with a cluster of normal ports allowing packets to be buffered at the shadow port in case the associated normal ports are sleeping. The Time Window Prediction (TWP) scheme is similar to the On/Off-1 algorithm in [7], wherein predictions are made about the number of packets traversing a port in a time window, and if this number is below a set threshold then the port is powered off. However, the difference in this scheme was that the sleep window is made adaptive by setting a bound on the increase in latency – if the latency of the buffered packets increases above a set bound, the sleep window is reactively reduced. The Power Save Mode (PSM) scheme was a modification to TWP to make it more aggressive for energy-savings by not considering traffic flows while making sleep decisions. This enabled higher energy savings at the cost of increased latency. The simulation results employing an enterprise network’s traffic patterns demonstrate a 18-21\% potential for energy savings using the TWP scheme. Their tradeoff study indicated that the wake-on-arrival capability and large buffer sizes could allow for a significant increase in energy-savings while ensuring minimal performance degradation. Chiaraviglio et al. [10] proposed an algorithm that creates an ordered list of nodes by decreasing power consumption, and for each element in the list, it first powers down the node and then checks if network-wide connectivity is maintained. If the network is disconnected, the node is powered back on and the algorithm proceeds to the next list element. Similar procedure is employed to power down links ensuring connectivity and maximum link load constraints are met. The simulation results on topologies similar to those of national ISPs confirmed the possibility to save more than 23\% of energy. The IEEE 802.3az – Energy Efficient Ethernet (EEE) standard was approved in September 2010 with the aim to reduce the energy consumption of computer networks [11]. While there was more than one scheme initially proposed for this standard, the approach of low-power idle, developed at Intel, was finally adopted [12]. Ethernet interfaces traditionally transmitted an auxiliary signal called IDLE when no data packets are transmitted to keep the transmitters and receivers synchronized. In 802.3az, in the case of no traffic, the link enters a low-power state and sends a Refresh signal over short durations allowing the link to consume less energy over large durations. Estimated savings by using EEE were projected to be $410 million/year in the USA, and over $1 billion/year globally. Herrería et al. enhanced the techniques of opportunistic sleeping in [25] by proposing four interface states – active, idle, transition, and sleep. The interface transitioned to sleep as soon as its buffer was empty, to allow for larger sleep durations, and the wake up back to active state was triggered either by a timer or the buffer length crossing a certain threshold. The simulation results showed a potential to save 75\% energy without noticeable impact to performance. Distributed algorithms to make sleeping decisions are presented in [80-82]. In [80], the approach involved selecting a router as the power saving router (PSR) by random election, and the PSR checking if network connectivity would be maintained if itself is disconnected from the network for a certain time period. If connectivity could be maintained, the PSR recomputed the routing table which would then be broadcast through the network. The numerical results from this approach indicated that up to 18\% of the power could be saved for the whole network. [81] utilized the periodic link state information (LSA) to decide which links to switch off in a distributed way. The proposed algorithm did not require the knowledge of the actual and past/future traffic matrices, and the results from realistic case studies indicated energy savings of up to 50\%. In [82], Patota et al. proposed DAFNES – a distributed algorithm for network energy saving based on stress-centrality. The stress centrality of a network node n refers to the number of shortest paths between any two endpoints which pass through node n. The proposed approach falls under the category of coordinated sleeping wherein switches, one at a time, made the decision to power off linecards independently, but the selection of the switch required synchronization by the exchange of control messages. The algorithm involved the computing and exchange of the stress centrality values between all switches, and the switch with the lowest stress centrality switched off its linecards in each iteration. The testing results showed the potential to save up to 50\% of network energy. [89] aimed to reduce the power consumption in data center networks by optimizing the number of SDN controllers active to serve the OpenFlow switches. The proposed heuristic first sorted all switches by decreasing order of the number of aggregated flow arrival rates, and then assigned each switch to a partially filled controller with the smallest but sufficient capacity. If no active controller was found with sufficient capacity, then a new controller node would be turned on. 3.2. Link rate adaptation The concept of link rate adaptation was first proposed by Christensen et al. [13] and was based on the realization that higher link speeds equated to higher energy consumption. The aim was to match link data rates with the traffic levels to make networks more energy proportional. In [14], Gunaratne et al. preliminarily explored rate adaptation for links between PCs and access-layer switches. The proposed algorithm made decisions to increase or decrease link rates based on the queue lengths on personal computer (PC) NICs and switch interfaces, and the simulation experiments on a university campus network confirmed that it was feasible to operate at lower data rates with no significant increase in delay. The work to develop a buffer threshold policy was extended in [15], [16] and [17]. In [15], a Markov model is developed for a state-dependent service rate, single-server queue while making the assumptions of standard Poisson arrival rates and exponential service rates. The proposed model is augmented by ensuring that service rate transitions occur only at service completions to replicate traditional Ethernet behavior. Single threshold policy, wherein service rates are switched to higher or lower rates if the output buffer utilization crosses a set threshold, and dual threshold policy, wherein upper and lower thresholds are defined and the service rate is switched to a higher rate when the output buffer utilization cross the upper threshold, and switched to a lower rate when the output buffer utilization falls below the lower threshold, are presented. The simulation experiments performed on traces from two university campuses indicated that links could be run at lower rates 99\% of the time with a 4 ms average increase in delay. To prevent frequent link rate changes adversely impacting performance in case of smooth and bursty traffic, current link utilization information is incorporated into the buffer threshold policy in [16]. An important insight in this work was that links should be operated at lower data rates for at most 50\% utilization, which equates to 5\% utilization at the higher rate, and thus, if the current link utilization exceeds 5\% at the higher rate, the link is not switched to the lower rate in order to balance performance with energy savings. An additional policy – time-out threshold policy – was proposed in [17] with the aim to reduce complexity of a NIC having link rata adaptation capability. The notion behind this scheme was to hard-set the time a link would run at the higher data rate and once this timer expired, and if the buffer utilization is below the set threshold, the link would switch to the lower data rate, and if not, it would continue to operate at the high rate. Transition from low to high data rates were immediate, triggered by the buffer utilization crossing the set threshold. Callegari et al. extended this work in [24] by incorporating transition times into their Markov model for a dual threshold policy, but their simulation experiments concluded that link rate adaptation was unsuitable for the NIC buffer sizes of the time. A mechanism to implement link rate adaptation was proposed in [18]. A MAC handshake protocol was presented wherein a MAC frame containing the desired line rate was exchanged between the two nodes comprising a link. A switch to a lower data rate initiated by one node was permitted only if the other node responded with an ACK; if the other node cannot switch to the lower data rate, it responded with a NACK, and the first node continued operation at its original rate. A switch to a higher data rate initiated by one node would always be responded with an ACK by the other node to ensure no performance degradation. This mechanism was formalized as Rapid PHY Selection (RPS) in [19] and was proposed as one of the candidate approaches for IEEE 802.3az. The emulation results indicated little to no impact, i.e. minimal increase in delay and no packet loss, on TCP and UDP file transfers using this scheme. Nedevschi et al. studied rate adaptation in [8] with the motivation being that operating links at slower speeds have twofold benefits – operating at slower frequencies consumes less energy, and operating at slower frequencies allows the use of dynamic voltage scaling (DVS) – making energy consumption scale quadratically with operating frequency. A practical algorithm is proposed which uses exponentially weighted moving average (EWMA) to make estimations about packet arrival times and uses current link utilization and operating rate information to ensure a rate change does not violate delay constraints. Two important conclusions from the simulation experiments performed on traces from Abilene were - (i) the granularity and distribution of operating rates played an important factor in the number of transitions and therefore the amount of energy saved; uniformly distributed link rates performed significantly better than exponentially distributed rates available in most networking hardware, and (ii) the time to transition between data rates also impacts the amount of energy savings, with higher transition times leading to reduced savings and higher delay. Mahadevan et al. [20] studied link state adaptation (LSA) and performed experiments by simulating a Web 2.0 workload in a production data center topology. The proposed strawman scheme reactively adapted link rates based on current utilizations, while the service level (SL) aware scheme added constraints to guarantee a minimum level of performance was maintained – ensuring the link utilization was below 70\% at any given time. The schemes assumed the existence of an Oracle that had perfect knowledge about upcoming traffic. The results showed that 16\% energy savings were possible employing the LSA scheme, while SL-aware LSA consumed slightly more energy but provided significantly better performance. The deployment considerations described two approaches – reactively making changes by collecting link utilization information using protocols such as SNMP, and proactively making changes by predicting future traffic patterns using simple models such as AR(1). However, the prediction-based approach led to over and under-estimations in their simulations. Abts et al. [21] leveraged the capability of modern plesiochronous links to operate in a dynamic range to make data centers more energy proportional. The intuition is that high-speed communication channels generally comprise of multiple links operating plesiochronously independent of each other and of the core router logic rate, and this independence is exploited to match operating rates with estimated bandwidth requirements. The proposed mechanism involved the switch tracking the amount of traffic traversing through it in a given time period, and if it exceeded a threshold defined for each link, the operating rate was doubled, and if it was less than the threshold, the operating rate was halved. The simulation …