Joint Server Selection and Routing for Geo-Replicated Services

Size: px
Start display at page:

Download "Joint Server Selection and Routing for Geo-Replicated Services"

Transcription

1 Joint Server Seletion and Routing for Geo-Repliated Servies Srinivas Narayana, Wenjie Jiang, Jennifer Rexford, Mung Chiang Prineton University Abstrat The performane and osts of geo-repliated online servies depend on whih data enters handle user requests, and whih wide-area paths arry traffi. To provide good performane at reasonable ost, servie providers adapt the mapping of user requests to data enters e.g., through DNS), and routing of responses bak to users i.e., through multi-homed route ontrol). Mapping and routing are typially managed independently, with mapping having limited visibility into routing deisions, response path latenies, and bandwidth osts. However, poor visibility and unoordinated deision-making an lead to worse performane and higher osts when ompared to a joint deision. In this paper, we argue that mapping and routing should ontinue to operate modularly, but ooperate towards servie-wide performane and ost goals. Our main ontribution is a distributed algorithm to steer ooperating, yet funtionally separate, mapping and routing provably towards a globally optimal operating point. Trae-based evaluations on an operational CDN show that the algorithm onverges to within 1% of optimum in 3-6 iterations. I. INTRODUCTION Online servie providers OSPs) arefully optimize the performane their ustomers experiene, sine even small inreases in user-pereived lateny an have signifiant impat on revenue [1]. To improve performane and reliability, OSPs run servies out of multiple geographially distributed servie loations e.g., data enters), eah typially peering with multiple ISPs. However, good performane must be balaned against operational osts resulting from the eletriity and bandwidth needed to serve large amounts of data on a daily basis [2], [3]. Client traffi loads, path latenies, and eletriity and bandwidth osts, vary over spae and time. To provide good performane at reasonable osts, OSPs adapt the wide-area paths arried by traffi in two key ways Fig. 1) mapping requests to speifi loations e.g., DNS rediretion [4], [5]), and routing responses bak to lients through peer seletion at the egress to the Internet e.g., multi-homed routing [2], [6]). These knobs allow an OSP to exploit path and ost diversity to adapt to time-varying demands, path performane, and osts. To our knowledge of the state of the art, mapping and routing are managed independently: mapping deisions are made without full visibility into the path metris, loads and deisions of the multi-homed) routing system. However, as we show in II, misaligned objetives, lak of information visibility and unoordinated deision-making an lead to bad performane and high osts. For example, the mapping may diret user requests to loations with limited upstream bandwidth, poor routing performane [4], [7], or expensive ISP onnetivity [2]. As a Google CDN study [4] shows, lients an experiene latenies inflated by several tens of milliseonds even when served by the geographially losest node, in part Fig. 1. Data Centers Multi-homed Internet onnetivity Internet Clients! IP Prefixes Mapping Nodes! DNS resolvers An online servie with multiple data enters aross the Internet. due to routing ineffiienies. In suh ases, better visibility into routing deisions and path performane an allow poorly performing lients to be mapped to other loations. A natural approah to improve the situation is to jointly ompute mapping and routing deisions after all, the OSP may have administrative ontrol over both. However, the ability to separately ompute deisions is desirable for two key reasons. First, an OSP may not be running its own infrastruture on all parts of its global traffi management system. For example, an OSP may employ an external mapping servie, e.g., [5], [8]. In suh ases, the OSP may only ditate highlevel poliies but may be unable to ompute or onfigure the mapping deisions diretly. Seond, a large OSP may already have operational mapping and routing infrastruture working independently [4], [7]. Hene, instead of building a monolithi system, we argue that mapping and routing should ontinue to operate modularly, but oordinate to ahieve the servie-wide performane and ost of a joint system. Our goal is to onstrut a distributed algorithm to oordinate administratively separate mapping and routing systems, and provably ahieve global performane and ost goals. Unlike prior works on OSP joint traffi engineering [2], [7], [9], this algorithm retains the modularity of mapping and routing while addressing major onerns that ouple their deisions e.g., link apaities). Our ontributions are as follows. Identifying hallenges in oordination. We show why a naive distributed algorithm, where mapping and routing iteratively optimize deisions even with full information visibility an lead to bad performane and high osts II). Distributed mapping and routing algorithm. We present a distributed algorithm that systematially addresses the auses of suboptimality, hene allowing modular mapping and routing to provably onverge to a global optimum IV). The solution is summarized in Fig. 2. Our optimization model aptures on-

2 Data enter Mapping nodes Data enter i ROUTING i Mapping deisions Client traffi volumes MAPPING... Data enter Routing deisions Path performane the senario in Fig. 4b), where a mapping initialization e.g., from geo-proximity) direts all traffi to the DC on the left. No traffi from this user reahes the other DC, so all routing deisions here are equally good. Suppose this DC hooses to route traffi through the 100ms path. Then, these mapping and routing deisions are loally optimal to eah other preventing the routing from exploring the globally optimal 50ms path. Fig. 2. Distributed mapping-routing algorithm IV). Data enter i solves loal problem ROUTING i to obtain multi-homed routing deisions, and mapping nodes solve MAPPING to ompute global mapping deisions. They then exhange measured traffi volumes, path performane, and omputed mapping and routing deisions, and iteratively reoptimize until onvergene. siderations like traffi demands, path performane, bandwidth osts, mapping poliies [5], [8], and link apaities III). Evaluation on CoralCDN traes. We evaluate the distributed algorithm on traes from CoralCDN [10], an operational ontent distribution network, and show that it onverges to within 1% of the global optimum in 3-6 iterations V). II. COORDINATING MAPPING & ROUTING Performing lient-mapping and network-routing independently an miss opportunities to improve performane or redue osts. Through an example OSP whih has two data enters DCs) eah with two ISP links, and one lient IP prefix, we highlight some hallenges that hinder independent mapping and routing systems from arriving at good olletive deisions. Misaligned objetives an lead to suboptimal deisions. Typially, mapping and routing systems work with different objetives e.g., mapping performs lateny and load-based DC seletion, while routing onsiders end-to-end performane and bandwidth osts to hoose a peer to forward responses. In Fig. 3a), the routing system at eah DC piks the peer with the least ost per unit bandwidth, while the mapping system piks the DC with the least propagation delay given urrent routing hoies resulting in a situation where the servie neither has globally optimal ost nor optimal lateny. Inomplete visibility an lead to suboptimal deisions. In Fig. 3b), the mapping and routing systems both optimize lateny, and users are sending traffi at the rate of 5Gb/s. Without information on link loads and apaities, the mapping system direts too muh traffi to the DC on the left, leading to large queuing delays and paket losses. If the mapping system uses information about link apaities, it ould easily diret most traffi to the alternate DC with ample spare apaity, and only slightly worse lateny. Coupled onstraints an lead to suboptimal equilibria. Even if mapping and routing have aligned objetives e.g., minimize lateny) and omplete visibility, optimizing their deisions separately taking turns an still lead to globally suboptimal situations. In Fig. 4a), mapping and routing are loally optimal given eah other s deisions, as traffi is served through the least lateny paths respeting link apaities. However, in the globally optimal alloation, all traffi is served by the DC on the right, using both peers. Bad routing deisions for prefixes with no traffi ontribution to a DC an lead to suboptimal equilibria. Consider $5/GB 40ms $1/GB 100ms a) $10/GB 90ms $12/GB 250ms Route-seleted Alternate route Map-seleted 100ms 1Gb/s 45ms 5Gb/s b) 10Gb/s 50ms 75ms Route-seleted Alternate route Map-seleted Fig. 3. a) Mapping is suboptimal beause of misaligned objetives; b) Mapping is suboptimal beause routing does not share link-related information. 4Gb/s 90ms 4Gb/s 75ms 4Gb/s 75% 25% a) 1Gb/s 50ms 3Gb/s 60ms Route-seleted Alternate route Map-seleted 100ms 2Gb/s 60ms 1Gb/s b) 2Gb/s 100ms 2Gb/s 50ms Route-seleted Alternate route Map-seleted Fig. 4. a) Separate mapping and routing an be ineffiient under oupled onstraints; b) Routing is suboptimal when mapping sends no traffi to a DC. III. OSP & OPTIMIZATION MODELS In this setion, we introdue the model for i) OSP mapping and routing deisions, ii) performane and ost goals, and iii) joint ost-performane optimization. A. Online Servie Provider Model Network. The OSP network model is summarized in Fig. 1. An online servie runs in a set I of servie loations i.e., data enters, ahes, replias, availability zones). Every loation i is onneted to the Internet through a set of ISP links, whih we denote by J i. Let C be the set of lients, where eah lient is an aggregate of real users in our experiments, these are IP prefixes). Clients an be direted to different loations depending on their proximity and the urrent load on the ompute infrastruture. Suh mapping of users to loations ours through a set of mapping nodes, suh as authoritative DNS servers or HTTP proxies. One a request is servied at loation i, the egress router piks one or more ISP-links j J i to send responses. The total traffi that a link j an support is limited by its apaity ap i j. Mapping Deisions hoie of servie loation). We denote α i as the proportion of traffi from lient mapped to loation i. We require that i α i = 1 for all, where α i [0,1]. Different users in the same IP prefix may be direted to different loations simultaneously. The mapping deisions are realized by a flexible mapping servie e.g., [5], [8], [11]. Eah user is served by its loal mapping node, e.g., a loal DNS server, and mapped to the designated servie loation. Routing Deisions hoie of egress ISP). Eah servie loation is onneted to the Internet by several links to different ISPs. We use the index j to denote a link that onnets a servie loation to an ISP. For every loation i, we denote by

3 β i j the proportion of traffi for lient served by loation i that is sent over link j J i, where j: j Ji β i j = 1. The responserouting deision only ontrols the first hop, and not the entire path. Typial routing deisions β i j are often integers {0,1}, sine one next-hop is hosen to reah eah lient. We relax this onstraint and allow OSPs to freely split lient traffi aross links. In pratie, frational routing an be realized by hash-based splitting or multi-homing agents [12]. Eah data enter is a stub autonomous system with no transit servie, hene routing hanges do not need BGP to reonverge. B. Performane and Cost Goals User-pereived lateny is a key performane metri for online servies, sine even small inrements an have signifiant effets on revenue [1]. User-pereived lateny an depend on round trip delays between users and data enters, proessing time within the ompute infrastruture, TCP dynamis, and even browser page loading time. In this paper, we fous on optimizing the round trip lateny for user requests, whih impats the ompletion time of short TCP flows [13], muh more than metris like bandwidth and paket loss. Average end-to-end lateny objetive. We use average endto-end path RTT ms/request) as our performane metri. Path latenies depend on whih loations handle requests, whih wide-area paths deliver traffi, and paket queuing and transmission delays along paths. A servie provider an obtain statistially averaged latenies through ative measurements [14] or lateny predition tools [15]. We use perf i j to denote the lateny from loation i to lient, when i piks link j J i to deliver traffi. Then the performane objetive is: perf = C vol α i β i j perf i j i I j J i ) /vol, where vol is the total request volume from lient. We ignore the impat of OSP traffi shifts on lateny as long as data enter-isp links operate safely under apaity. This is a standard simplifiation used in prior works [2], [6]. Average ost per request objetive. Operational osts are an important onsideration for OSPs [2], [3]. In this paper we fous on ISP bandwidth osts, whih an be signifiant on its own for large OSPs [2]. For tratability, we assume a ost funtion whih is linear on the amount of data sent. The ost metri is average ost per request $/request): ost = C vol α i β i j prie i j i I j J i ) /vol where prie i j is the ost per request on link j J i of data enter i. In pratie ISPs employ omplex priing funtions e.g., 95th perentile harging). However optimizing a linear ost on every harging interval an redue monthly 95th perentile osts [2]. C. Joint Cost-Performane Optimization The goals of minimizing lateny and minimizing osts an often be at odds. For example, an ISP with riher Internet peering may offer lower latenies to more lients, albeit at higher ost, than its ompetitors [2]. To strike a balane between ost and performane, we introdue a weight K that reflets the amount of additional osts $/request) that an OSP is willing to pay for one unit of performane improvement ms/request). Hene, the OSP s goal is to minimize ost + K perf, whih aptures pareto-optimal tradeoffs between perf and ost. For additional flexibility, we allow OSPs to express mapping poliies in terms of traffi split ratios w i with tolerane ε i ) or absolute request aps B i per data enter i. This allows an OSP to balane load between data enters in a weighted round robin fashion [8] or limit load at ertain sites [5]. We formulate the joint optimization problem as follows: GLOBAL minimize ost + K perf 1a) subjet to vol α i β i j ap i j, j 1b),i: j J i vol α i w i ε i, i 1) vol vol α i B i, i 1d) variables i α i = 1, β i j = 1, i, j J i α i 0, i,, β i j 0, i, j, 1e) 1f) where 1b) ensures the aggregate link traffi does not exeed apaity [2], [5], [9], [16]. Constraints 1) and 1d) enfore traffi split ratios and bandwidth aps respetively. Constraints 1e) and 1f) ensure that proportions sum up to one. The problem 1) is a linear program on α and β separately, but non-onvex when both are variables. As suh, it annot be solved using standard onvex programming. However, it an be onverted into an equivalent LP whose variables ouple mapping and routing) and solved effiiently Appendix A). IV. OPTIMAL DISTRIBUTED ALGORITHM We design effetive ways to avoid suboptimal equilibria IV-A) and present an optimal distributed algorithm IV-B). A. Avoiding Suboptimal Equilibria A straightforward distributed solution to the joint problem 1) is to alternate between the mapping nodes optimizing over α given routing deisions β), and the routing nodes optimizing over β given mapping deisions α). However, suh separate optimizations often lead to loal optima II). We systematially address the auses of suboptimality from II. First, we align mapping and routing objetives expliitly by onstruting both to solve the joint problem 1) iteratively. Further, we mandate that eah system has visibility into information needed to solve the joint problem, by requiring speifi loal measurements and exhange of statistis IV-B). Next, we illustrate how we mitigate the oupled onstraints problem through an example network in Fig. 5 a). Let α and β be the mapping and routing deisions respetively. If the lient has one unit of demand, optimal lateny is ahieved when α = 1,β = 1/4. However, by separate optimizations, we an end up in loal optima, e.g., α = 1/2,β = 1/2, whih are valid but globally suboptimal mapping and routing profiles.

4 ap=100 ap= ms 60ms ap=3/4 1-β 1-α α 75ms 50ms ap=1/4 β Available route a) Example network b) GLOBAL ) GLOBAL Φ Fig. 5. An example of loal optima with separate optimization. Fig. 5 b) visualizes how the loal optimum is reahed from an initial starting point. The olor-shaded region represents the feasible deision spae. The urvy boundaries reflet the apaity onstraints of the 50ms and 60ms links. The olor oding means the objetive value, i.e., lateny, where red is high and blue is low. All points on these boundaries exept at α = 1 are loal optima, whih an be very ineffiient. In general, the feasible spae determined by link apaity and mapping poliy onstraints is non-onvex. Motivated by the observation that an ill-shaped feasible region does not enable effiient distributed optimization, we aim to work around the boundaries implied by the hard apaity onstraints. We show later that these are in fat the only onstraints leading to suboptimal equilibria in Appendix B.) We relax the link apaity onstraint by introduing a penalty funtion. In partiular, we utilize a piee-wise linear funtion Φ i j ) often used in ISP traffi engineering [17] to apture osts due to high link utilization. The penalty term Φ improves the predition of lient performane by apturing the effets of link ongestion on queuing delay. In the refined global optimization, namely GLOBAL Φ, we replae the objetive 1a) by obj g α,β) = ) α i β i j vol priei j + K perf i j i j + i j vol Φ i j α i β i j vol )/ J, 2) whih an be viewed as a joint ost-performane optimization with link ongestion onsideration. The onstraint 1b) that aused the suboptimal equilibria is now removed from the problem. Figure 5) shows that the solution of this refined problem formulation onverges to the global optimum by alternately fixing one dimension and optimizing the other. The whole deision spae is now feasible, but the violation of link apaity will inur a high ost. Note that the global optima shown in Figures 5b) and 5) are different, as the ongestion ost is not onsidered in the original formulation. B. Distributed Mapping and Routing We present a distributed solution that builds on the existing pratie of omputing mapping and routing deisions independently. The key idea is to design optimization problems for eah system, i.e., mapping and routing, in whih i) they have aligned objetives, and ii) eah system omputes only its own loal) deision variables, using loal onstraints and exhange of appropriate statistis. In partiular, mapping nodes olletively solve the following mapping problem: MAPPINGβ) minimize obj g 3a) α i vol subjet to w i ε vol i, i 3b) variables i α i vol B i, i α i = 1,, α i 0, i, Further, edge routers together solve the following problem: 3) 3d) ROUTINGα) minimize obj g 4a) subjet to variables β i j = 1, i, j J i β i j 0, i, j, 4b) Note that ROUTING does need to know about MAPPING s load balaning poliies. Further, MAPPING and ROUTING are both onvex optimizations, whih an be effiiently solved. A distributed solution, the alternate projetion algorithm, proeeds as follows: given routing deisions β as input, mapping nodes solve 3) and optimize over α, and given mapping deisions α as input, edge routers solve 4) and optimize over β. The two steps are arried out iteratively until solutions onverge. We allow mapping and routing problems to be solved at different time-sales, with the only requirement that 3) does not start until 4) is fully solved, and vie versa. This is easily ensured by having eah omponent wait to reeive inputs from the other before solving its own loal optimization problem. However, we showed in Figure 4b), in general the alternate projetion algorithm may still lead to suboptimal equilibria, when some lients send no traffi to a data enter. To mitigate

5 this no traffi problem, we introdue a refinement to routing in addition to the optimality of 4). We later prove the optimality of suh a refined routing strategy in Appendix B.) In pratial omputation of routing deisions, we introdue an approximation to 4) by inrementing an infinitesimally small demand to a lient-server pair with zero traffi, i.e., α i δ if α i = 0, where δ is a small positive onstant [18]. This approximation is often used to ease the omputation so that standard optimization tehniques an be applied. However, lients do not need to send real traffi, making this approah pratially appealing. We show that the alternate projetion algorithm with the refinements introdued above provably onverges to the global optimum of GLOBAL Φ : Theorem 1: Alternate projetions of MAPPING and ROUTING onverge to an optimal solution of GLOBAL Φ. Note that ROUTING an be further deomposed into loal versions at eah data enter, as both its objetives and onstraints an be deoupled, e.g., obj g = i obj i) g, where obj i) g = ) α i β i j vol priei j + K perf i j / vol j J i, + Φ i j α i β i j vol )/ J j J i is the loal objetive funtion at data enter i. There is no need for oordination among data enters, and their deisions olletively attain the optimal solution to 4). The mapping problem 3) an also be distributed among mapping nodes; indeed [5] studies this problem. We omit the details here. The distributed algorithm is summarized in Fig. 2. Mapping nodes olletively measure lient request rates, and solve the global mapping problem MAPPING using routing deisions and path latenies from the routing system. Eah data enter loally measures path latenies to lients, and solves ROUTING i using mapping deisions and traffi volumes from the mapping system. We assume that other optimization parameters namely onfiguration of data enters, peering links, link apaities and mapping poliies vary slower than the deision reoptimization time, and are globally up to date. C. Optimality of distributed solution Due to spae limitations, we present a proof sketh and defer the formal proof to the Appendix. The proof proeeds in three steps. We first show that there exists a equilibrium point α,β ), suh that α is a best response to β, and vie versa. We define the best response as an optimal projetion Def. 2), and the resulting equilibrium as a Nash equilibrium Def. 3). We further prove that α,β ) is also an optimal solution to GLOBAL Φ, whih is one of our main ontributions and signifiantly generalizes previous results [18]. We finally show that the alternate projetion algorithm leads to a Nash equilibrium point, whih ompletes the proof. The rest of this setion introdues these formalities to assist our proof. First, we need a way to apture the marginal ost f i j of serving lient from data enter i over link j: Definition 1: Let the metri f i j be defined as ) f i j α i,β i j ) = vol priei j + K perf i j / vol +vol Φ j α i β i j vol )/ J We next define the optimal projetion of mapping and routing strategies, respetively: Definition 2: i) Given fixed routing deisions β, a set of mapping deisions α is alled an optimal projetion onto mapping spae if α is an optimal solution to 3). ii) Given fixed mapping deisions α, a set of routing deisions β is alled an optimal projetion onto routing spae if, i,, we have f i j α i,βi j ) f i j α i,β j ) for all j, j J i suh that β j > 0. Lemma 1: If β is an optimal projetion given α, then β is also an optimal solution to 4). It is easy to hek that an optimal solution to 4) is not neessarily an optimal projetion. We introdued an approximation in IV-B by inrementing small traffi demand to every lient-server pair. We then an alulate the optimal projetion of routing by solving refined 3) using standard optimization tehniques. We next define the notion of Nash equilibria. Definition 3: A set of mapping and routing deisions α,β ) is alled a Nash equilibrium if α is an optimal projetion given β, and β is an optimal projetion given α. V. DISTRIBUTED SOLUTION EVALUATION Experiment setup. CoralCDN [10] is a ahing and ontent distribution platform running on Planetlab. Our request-level trae olleted at 229 Planetlab sites running Coral onsists of over 27 million requests and a terabyte of data, orresponding to Marh 31st, To emulate multi-homed servie deployment, we luster Planetlab sites in approximately the same metropolitan area and treat them as ISPs peering with the same servie loation similar to the approah followed in [12]. Our setup has 12 servie loations distributed in North Ameria, Europe and Asia with 3-6 ISP onnetions eah. Requests arrive from about IP prefixes. For latenies to these prefixes, we use iplane [14], a system that ollets wide-area network statistis to Internet destinations from Planetlab vantage points. To orrelate the traffi and lateny data, we narrow down our lient set to IP prefixes, whih ontribute 38% traffi by requests) of the entire trae. The osts per unit bandwidth for ISP links are derived from a reallife loud bandwidth priing plan [19]. We refer the reader to Appendix D for full details. Benefits of joint optimization. We evaluate the benefits of full information visibility for ost and perf in isolation. Coneptually, this orresponds to setting K = 0 and K = respetively in the GLOBAL optimization. For perf, we ompare against a baseline mapping that optimizes for average lateny assuming per-lient least lateny path is used to reah eah lient from eah data enter, and a routing that optimizes perf. From hourly optimizations on the CDN trae, we find that GLOBAL ahieves 10ms smaller perf than this baseline on average.

6 Adaptive Cold start # iterations # instanes auray % TABLE I. SOLUTION OPTIMALITY AND CONVERGENCE. For ost, we use a baseline mapping that optimizes for average ost assuming average per-link ost is inurred at eah data enter, and routing that optimizes ost. GLOBAL ahieves between 3-55% lower ost aross the trae. Optimality to within 1% in 3-6 iterations. Over 24 hourly problem instanes through the trae, we reord the number of iterations for onvergene of mapping-routing alternate projetions, and the differene in %) of the onverged objetive from the optimal value of GLOBAL Φ, i.e., auray. As a onvergene riterion, we test if objetive values from suessive iterations are within a fixed perentage say 1%) of eah other. We examine two sets of initial onditions, namely old start, i.e., mapping is initialized with uniform round robin splitting, and adaptive, i.e., mapping starts from the onverged setting of the previous hour. The results are summarized in Table I. We find that i) onvergene ours in 3-6 iterations, ii) adaptive mapping generally onverges faster than old start, and iii) the onverged objetives are indeed within the fixed tolerane 1%) of the GLOBAL Φ optimum. Optimization runtime under 1 minute on a modern mahine. A single iteration of the alterate projetion algorithm with 12 loations, 49 links and lients takes about 30 seonds on a desktop mahine with a 2.40GHz two-ore CPU. Impat of penalty funtion Φ on lateny. We evaluate how the value of perf is inreased when the distributed algorithm optimizes a lateny objetive with link utilization penalties obj g, eqn. 2)), instead of optimizing perf diretly. Note that this relaxation in general auses an inrease in perf. Over hourly instanes through the trae, we find a uniform inrease of 10ms perf over optimizing perf diretly using 100% available link apaities. The gaps are 8ms and 5ms respetively when using only 95% and 90% of link apaities. Piking a ost-performane tradeoff. The GLOBAL optimization trades off performane for ost through a parameter K. However, it an be hallenging to know what a good tradeoff looks like. To understand this better, we visualize a pareto-optimal tradeoff urve between ost and perf for the the day in Fig. 6, as K varies. We see that there is a rih tradeoff spae between ost and lateny, with perf ranging between ms 88% inrease possible from optimal perf) and ost ranging between $/GB 53% inrease possible). A servie provider ould manually pik a sweet spot from this urve for example, the value of K orresponding to the knee at perf 98ms and ost $0.175/GB. VI. RELATED WORK There is a rih literature exploring the benefits of visibility between networks, appliations and servie infrastruture. Provider networks and selfish appliations. P4P [20] and IETF ALTO [21] advoate that appliations should use hints from ISPs to improve both their performane and ISP ontrol Fig. 6. ost $$/GB) perf ms) Pareto-optimal tradeoff urve between lateny perf) and ost $/GB). over their traffi. OSP traffi engineering differs from this setting sine the OSP ontrols the user-replia mapping diretly. ISP and CDN ollaboration. There have been many reent proposals suggesting ollaboration between ISPs and ahing infrastruture to improve servie performane [22] [24]. In ontrast, OSP traffi engineering onsiders paths aross ISPs onneted from distributed servie loations. This hanges the problem beause flexible ontent plaement inside a single ISP an provide signifiant performane gains. However, repliation aross ISPs an be expensive for OSPs hosting dynami ontent [7], [25]. Our optimization model more losely resembles earlier proposals to mitigate bad interation between replia seletion and ISP traffi engineering given the set of replias [16], [18], [26]. However, these works onsider routing within a single administrative domain, and do not inorporate flexible mapping poliy onstraints, e.g., 1). OSP joint traffi engineering. Unlike Entat [2], our algorithm ahieves joint system gains by oordinating modular request-mapping and response-routing systems. Our goals are similar to those of PECAN [7] but our model also onsiders aspets whih intimately ouple mapping and routing deisions e.g., link apaities, bandwidth osts). Xu et al. [9] distribute the joint optimization for sale; however their separation of variables does not deouple the mapping and routing systems. VII. CONCLUSION We address the problem of jointly optimizing requestmapping and response-routing for online servies, while retaining the funtional separation between them. We highlight the hallenges in ahieving the gains of a joint system in a distributed setting, and mitigate these hallenges through a distributed algorithm that provably onverges to a globally optimum point. Hene, OSPs an attain the benefits of a joint system namely better performane at lower ost while only slightly refining the behavior of existing infrastruture. Aknowledgment. This work was funded by NSF grant , DARPA grant FA C-0254, and a gift from HP. We thank L. Vanbever, B. Balasubramanian, M. Wang and W. Lloyd for feedbak and H. Madhyastha for iplane traes. REFERENCES [1] J. Hamilton, The ost of lateny, /10/31/TheCostOfLateny.aspx. [2] Z. Zhang, M. Zhang, A. Greenberg, Y. C. Hu, R. Mahajan, and B. Christian, Optimizing ost and performane in online servie provider networks, in Pro. NSDI, 2010.

7 [3] A. Qureshi, R. Weber, H. Balakrishnan, J. Guttag, and B. Maggs, Cutting the eletri bill for internet-sale systems, in Pro. ACM SIGCOMM, [4] R. Krishnan, H. V. Madhyastha, S. Srinivasan, S. Jain, A. Krishnamurthy, T. Anderson, and J. Gao, Moving beyond end-to-end path information to optimize CDN performane, in Pro. IMC, [5] P. Wendell, J. W. Jiang, M. J. Freedman, and J. Rexford, DONAR: Deentralized server seletion for loud servies, in Pro. ACM SIG- COMM, [6] D. K. Goldenberg, L. Qiu, H. Xie, Y. R. Yang, and Y. Zhang, Optimizing ost and performane for multihoming, in Pro. ACM SIGCOMM, [7] V. Valanius, B. Ravi, N. Feamster, and A. C. Snoeren, Quantifying the benefits of joint ontent and network routing, in Pro. ACM SIGMETRICS, [8] Amazon Route 53, [9] H. Xu and B. Li, Joint request mapping and response routing for geodistributed loud servies, in Pro. IEEE INFOCOM, [10] M. J. Freedman, E. Freudenthal, and D. Mazières, Demoratizing ontent publiation with Coral, in Pro. NSDI, [11] Y. Zhu, B. Helsley, J. Rexford, A. Siganporia, and S. Srinivasan, LatLong: Diagnosing wide-area lateny hanges for CDNs, IEEE Transations on Network and Servie Management, vol. 9, [12] A. Akella, B. Maggs, S. Seshan, A. Shaikh, and R. Sitaraman, A measurement-based analysis of multihoming, in Pro. ACM SIG- COMM, [13] N. Cardwell, S. Savage, and T. Anderson, Modeling TCP lateny, in Pro. IEEE INFOCOM, [14] H. V. Madhyastha, T. Isdal, M. Piatek, C. Dixon, T. Anderson, A. Krishnamurthy, and A. Venkataramani, iplane: An information plane for distributed servies, in Pro. OSDI, Nov [15] M. Szymaniak, D. Presotto, G. Pierre, and M. van Steen, Pratial large-sale lateny estimation, Computer Networks, vol. 52, pp , May [16] J. W. Jiang, R. Zhang-Shen, J. Rexford, and M. Chiang, Cooperative ontent distribution and traffi engineering in an ISP network, in Pro. ACM SIGMETRICS, [17] B. Fortz, J. Rexford, and M. Thorup, Traffi engineering with traditional IP routing protools, IEEE Communiation Magazine, [18] D. DiPalantino and R. Johari, Traffi engineering vs. ontent distribution: A game theoreti perspetive, in Pro. IEEE INFOCOM, [19] Amazon EC2 Priing, [20] H. Xie, Y. R. Yang, A. Krishnamurthy, Y. G. Liu, and A. Silbershatz, P4P: Provider portal for appliations, in SIGCOMM, [21] IETF ALTO, [22] I. Poese, B. Frank, G. Smaragdakis, S. Uhlig, A. Feldmann, and B. Maggs, Enabling ontent-aware traffi engineering, ACM SIG- COMM Computer Communiation Review, [23] A. Sharma, A. Venkataramani, and R. K. Sitaraman, Distributing ontent simplifies ISP traffi engineering, in ACM SIGMETRICS, [24] B. Frank, I. Poese, Y. Lin, G. Smaragdakis, A. Feldmann, B. Maggs, J. Rake, S. Uhlig, and R. Weber, Pushing CDN-ISP Collaboration to the Limit, SIGCOMM CCR, July [25] S. Agarwal, J. Dunagan, N. Jain, S. Saroiu, A. Wolman, and H. Bhogan, Volley: automated data plaement for geo-distributed loud servies, in Pro. NSDI, [26] I. Poese, B. Frank, B. Ager, G. Smaragdakis, and A. Feldmann, Improving ontent delivery using provider-aided distane information, in IMC, APPENDIX A A CENTRALIZED APPROACH TO THE JOINT PROBLEM We present an equivalent onvex formulation of the joint mapping and routing optimization problem that aepts an effiient entralized solution. GLOBAL is a non-onvex optimization problem on variables α i,β j ), and hene annot be solved using standard onvex programming tehniques. It is not lear whether loal optima exist and how bad they are relative to the global optimum. However, we show that 1) an be translated into an equivalent onvex optimization problem by introduing a new set of variables X i j, whih represents the fration of lient traffi that is mapped to data enter i through link j. There is a diret onnetion between the two sets of variables: X i j = α i β j, i, j,) We an rewrite 1) into the following onvex program: GLOBAL X minimize subjet to variables ) X i j vol priei j + K perf i j i j, j Ji vol X i j vol vol w i ε i, i 5a) 5b), j J i vol X i j B i, i 5),i: j J i vol X i j ap j, j X i j = 1, i j X i j 0 i, j, 5d) 5e) The size of the linear program is O J C ). The feasible sets of GLOBAL and GLOBAL X are relevant, and the two problems have the same optimal objetives. Theorem 2: GLOBAL and GLOBAL X are equivalent. They have the same optimal objetive values. Further, their optimal solutions an be derived from eah other. Proof: First, the optimal value of GLOBAL X is a lowerbound on that of GLOBAL. Consider any feasible solution {α i,β i j } of 1). We onstrut X i j = α i β i j, whih is also feasible for problem 5). It an be readily verified that two solutions ahieve the same objetive value. Seond, the optimal value of GLOBAL is a lower-bound on that of GLOBAL X. Consider any feasible solution {X i j } of 5). We onstrut α i = j Ji X i j, β j = X i j /α i if α i > 0 and β j = 1/ J i otherwise. It an be verified that {α i,β i j } is feasible for GLOBAL, and ahieves the same objetive value as GLOBAL X. We establish a one-to-one mapping between variables of 1) and 5), and hene they obtain the same optimal objetives and solutions. The above theorem shows a simple approah to solving the joint problem by deploying a entralized oordinator. After olleting all neessary information suh as {perf i j,prie i j,vol }, we rewrite 1) into 5), whih an be solved using standard linear programming tehniques. We then reonstrut {α i,β j } from {X i j } using the above methods and distribute them to all mapping nodes and egress routers. APPENDIX B OPTIMALITY OF DISTRIBUTED MAPPING AND ROUTING A. Proof of Lemma 1 Proof: To show that β is an optimal solution to the routing problem 4), we hek the KKT onditions for 4). These onditions hold if and only if α i f i j α i f i j, j, j J i

8 and β i j > 0. This is true beause when α i > 0, the definition of optimal projetion implies the above ondition. When α i = 0, the optimality ondition holds too. B. Existene of Nash Equilibrium We follow the definitions and proofs of onvergene and optimality in [18]. However, we annot diretly apply their results beause of the presene of data enter load balaning onstraints 3b) and 3). Theorem 3: There exists a Nash equilibrium. Proof: We show this by onstrution. Let X be an optimal solution to GLOBAL X with refined objetive funtion obj g and removing apaity onstraint 1b). We onstrut a set of mapping and routing deisions as follows: αi = j Ji Xi j, β i j = X i j /α i if α i > 0. If α i = 0, we set β j = 1 for some j = argmin j Ji f i j, and set β j = 0 for j J i\{ j }. By Theorem 2, it is easy to hek that α,β ) is also an optimal solution to GLOBAL. Therefore, α must be an optimal solution to MAPPINGβ ), as otherwise we an find a better α to improve the global objetive funtion. By definition, α is an optimal projetion given β. Similarly, β is an optimal solution to ROUTINGα ). The KKT onditions for 4) imply that the definition of optimal projetion holds when αi > 0. When αi = 0, our hoies of β j stritly follows the optimal projetion definition. Combining the two ases shows β is an optimal projetion given α. Therefore, α,β ) is a Nash equilibrium. C. Optimality of Nash Equilibria Theorem 4: Let α,β ) be a Nash equilibrium. Then {α,β } is an optimal solution to GLOBAL. Proof: Construt a set of variables Xi j = α i β i j for i, j,). We show that {Xi j } is an optimal solution to GLOBAL X, given that {αi,β i j } is a Nash equilibrium. It is easy to hek that {Xi j } are feasible for GLOBAL X. Then, it suffies to show that the KKT onditions for GLOBAL X hold with our hoie of {Xi j } and appropriate hoies of other parameter, e.g., Lagrange multipliers, involved in the optimality onditions whih we disuss shortly). In a slight abuse of notation below, we write f i j X) = vol prie i j + K perf i j )/ vol +Φ j X i j vol ) vol / J. Here we onsider the mapping onstraint 3b) only and 3) should follow similarly. We write down the KKT onditions for GLOBAL X. There exist dual optimal variables {λ i1 0,λ i2 0} i suh that f i j + λ i1 λ i2 ) g f i ) j + λ i 1 λ i 2) g, if X i j > 0, λ i1 X i j g w i ε i = 0, i j ) λ i2 X i j g w i + ε i = 0, i j where g = vol / vol. The first ondition originates from the stationarity requirement, and an be interpreted as follows: when a link j J i link j of data enter i) is used to reah lient, e.g., X i j > 0, its assoiated marginal ost must be no greater than the marginal ost of any other link j J. Note that the marginal ost onditions are similar to those in [18], but slightly more ompliated due to the presene of dual variables and onstraints. The seond and third optimality onditions are omplementary slakness for the dual variables and orresponding onstraints. Next we onsider a Nash equilibrium α,β ). We an write down the KKT optimality onditions for MAPPING and ROUTING, respetively, beause the definition of optimal projetion implies the optimality to the two optimization problems. Following the same notations, we have: i) KKT optimality onditions for MAPPING: there exist dual optimal variables {µ i1 0, µ i2 0} i suh that βi j f i j + µ i1 µ i2 )g j J i β j f i j + µ i 1 µ i 2)g, if αi > 0, j J i ) 7) µ i1 αig w i ε i = 0, i 6) µ i2 ) αig w i + ε i = 0, i Similarly, the first ondition is the stationarity requirement, and the seond and third onditions are the omplementary slakness for the dual variables. ii) KKT optimality onditions for ROUTING: f i j f i j for j, j J i and β i j > 0, i,) where α i > 0 8) It follows that the values of f i j are equal for all j suh that β i j > 0 and α i > 0. By our onstrution, Xi j = α i β i j. We next show that {Xi j } satisfy the optimality onditions 6), given the KKT onditions 7)-8) established for {αi,β i j }. Without loss of generality, onsider Xi j > 0 for some i, j,) and j J i, and we have αi > 0,β i j > 0. From ROUTING optimality onditions 8), we know that f i j takes the same value for all j J i when β > 0. By our j hoie of βi j > 0, we have j J i β j f i j = f i j j J i β j = f i j 9) Consider any link j in data enter i, i.e., j J i. There exists at least one link j J i suh that β j > 0. Applying the optimality onditions 8) on i, j,), we have f i j f i j 10) Further, by an argument similar to 9), we have f i j = β j f i j 11) j J i To show that 6) holds for {Xi j }, it suffies to find a set of parameters {λ i1 0,λ i2 0} for all i suh that those equalities and inequalities in 6) hold. We laim that the dual

9 optimal variables {µ i1, µ i2 } in 7) an be used as a hoie of {λ i1,λ i2 }. We then show that the KKT optimality onditions 6) are satisfied with suh hoies. Consider Xi j > 0 for some i, j,) and j J i, we write down its assoiated marginal ost f i j, and ompare it against that of any other link j J i for the same lient. We have f i j + λ i1 λ i2 ) g = f i j + µ i1 µ i2 ) g hoie of dual variables) = βi j f i j + µ i1 µ i2 ) g j J i by 9) β j f i j + µ i 1 µ i 2) g by 7) j J i = f i j + µ i 1 µ i 2) g by 11) f i j + µ i 1 µ i 2) g by 10) = f i j + λ i 1 λ i 2) g hoie of dual variables) So we arrive at the onlusion that {Xi j } are optimal solutions to GLOBAL X. Therefore, {αi,β i j } are also optimal solutions to GLOBAL as they attain the same optimal objetive values. D. Convergene to Nash Equilibrium Theorem 5: Alternate projetion of MAPPING and ROUTING onverges to a Nash equilibrium. Proof: Consider the refined funtion obj g 2, whih is the ommon objetive for both MAPPING and ROUTING problems. By the definition of optimal projetions, the objetive value of obj g is a dereasing sequene in the alternate projetion algorithm. In addition, the objetive value is lowerbounded by the optimal value of GLOBAL X. Therefore, there exists a limit point of the sequene. Similarly, we an argue that the α, β) sequene also has a limit point, sine the the feasible spae of {α,β} is ompat and ontinuous. It is not diffiult to hek that the limit point must be be a Nash equilibrium, as otherwise we an find another feasible solution that is within an ε-ball of the the limit point, whih gives a lower objetive value than the limit of obj g, whih ontradits our assumption at the beginning. In our evaluations, we sale this funtion suh that 100% utilization gives a 250ms queuing lateny, orresponding to some standard router buffer sizes. APPENDIX D EXPERIMENT SETUP We evaluate the benefits of the joint mapping-routing approah using trae-based simulations for CoralCDN [10], a ahing and ontent distribution platform running on top of Planetlab. Planetlabis a servie deployment platform whih spans more than 800 sites geographially distributed in six ontinents. We orrelate the traffi data with lateny measurements from iplane [14] whih ollets various wide-area network statistis to destinations on the Internet from a few hundred vantage points hosted on Planetlab nodes). All data orresponds to Marh 31st, Emulating data enters by lustering Coral nodes: Most Planetlab sites are single-homed i.e., have just one ISP onneting them to the Internet. Therefore, it is hallenging to demonstrate the benefits of our sheme on Coral, beause the only deision that ontrols the lateny and ost of user requests is the mapping deision. Instead, we adopt the approah followed in [12] where we treat multiple Planetlab sites in the same general metropolitan area as different egress links of a single data enter loated in that area. The density of Planetlab sites around ertain ities makes suh lustering of sites into data enters possible. We handpiked 12 data enters loated in North Ameria, Europe, Asia and South Ameria, with a minimum of three Coral nodes in eah data enter and a maximum of six. Fig. 7. We show the loations of the data enters in Figure 7. Loation of the 12 emulated data enters. APPENDIX C PERFORMANCE PENALTY FUNCTION ISP traffi engineering models link ongestion ost with a onvex inreasing funtion Φr j ) on the link traffi load, i.e., r j. The exat shape of the funtion is not important, and we use the same pieewise linear ost funtion as in [17], given below: r j 0 r j /ap j < 1/3 3r j 2/3 ap j 1/3 r j /ap j < 2/3 10r j 16/3 ap j 2/3 r j /ap j < 9/10 Φ j r j,ap j ) = 70r j 178/3 ap j 9/10 r j /ap j < 1 500r j 1468/3 ap j 1 r j /ap j < 11/ r j 16318/3 ap j 11/10 r j /ap j < Routable prefixes as lient aggregates: We aggregate users into routable Internet prefixes from a RouteViewsFIB dump), for two reasons. First, routing deisions at data enter egress routers happen at this granularity, hene it is a natural hoie. Seond, the aggregation of IP prefixes in routing tables enables us to deide on reahability and lateny information for a larger number of lient IPs from the same set of initial lateny measurements. The number of routable IP prefixes is large 350,000) but only of these send any traffi to Coral. Traffi data from CoralCDN: We use a worldwide request trae from CoralCDN whih lists a request timestamp, the Planetlab site at whih it was reeived, a lient IP address, and the number of bytes involved in the transfer. Overall, this trae ontains 27 million requests and over a terabyte of data.

10 For our experiments, we aggregate the requests made by lients into their most speifi IP prefix from the RouteViews dump, and average their request rates for eah hour. Lateny measurements from iplane: We extrat round trip lateny in milliseonds) from the iplane logs, whih ontain traeroutes made to a large number of IP addresses from various vantage points at Planetlab sites. We only use lateny information from vantage points whih are also egress links in the data enters we piked in Figure 7, whih limits us to 49 vantage points. We assign the lateny to eah destination IP address to the orresponding most-speifi IP prefix from the RouteViews dump, assuming that the lateny is representative. If there is lateny data for multiple IP addresses from the same prefix, we use their average value. The logs provide only one set of lateny values for the entire day, and we assume that the paths are lightly loaded when the probes our hene treating these measured latenies as path propagation delays. Piking a set of workable lient prefixes: The set of destination IP prefixes for whih lateny data is available is not uniform aross vantage points. Hene, we only retain data for destinations whih are reahable from at least one vantage point belonging to eah data enter. This allows us to perform mapping of any lient to any data enter. Next, we interset this set of destinations with those lient prefixes whih send traffi to Coral CDN at any time through the day. We are left with a set of about prefixes, whih onstitute about 39% of the total traffi trae by requests and 38% by bytes. Capaity estimations: The apaities of links onneting Planetlab sites to the Internet are not publily available, and even if they are, Planetlab mahines are shared aross multiple servies eah possibly with bandwidth aps. Instead of attempting to determine apaity or bandwidth aps from ground truth this information is not publily available to our knowledge), we estimate them through traffi volumes in the trae. The key assumption we make for setting apaities from observed traffi is that links are provisioned for a speifi peak utilization. We determine the peak request rates reeived by Coral nodes whih are part of the emulated data enters. We sale these numbers as follows: 1) we aount for a speifi peak utilization 80%), hene saling by 100/80; 2) we aount for the traffi reeived by all Coral nodes inluding those not part of the emulated data enters) through saling by the ratio of peak hourly traffi over all Coral nodes to the peak hourly traffi over the hosen Coral nodes. Bandwidth ost estimation: We assume a simple harging model based on total amount of data transferred over a given period of time. We employ a performane-based model where we assign a higher ost to a link whih provides lower lateny paths to a higher proportion of traffi volume as determined from the traffi and lateny trae information). To quantify this, we first determine the number of requests servied by eah Coral node assuming that lients are servied by the latenywise losest Coral node. We then use table II motivated by prie values from Amazon EC2 bandwidth priing [19]) to determine the harging prie per GB of data sent over that link. Data Center Popularity total request volume per day: 10 5 ) TABLE II. Priing $ per GB) PERFORMANCE-BASED PRICING MODEL.

Joint Server Selection and Routing for Geo-Replicated Services

Joint Server Selection and Routing for Geo-Replicated Services Joint Server Selection and Routing for Geo-Replicated Services Srinivas Narayana Joe Wenjie Jiang, Jennifer Rexford and Mung Chiang Princeton University 1 Large-scale online services Search, shopping,

More information

Fast Distribution of Replicated Content to Multi- Homed Clients Mohammad Malli Arab Open University, Beirut, Lebanon

Fast Distribution of Replicated Content to Multi- Homed Clients Mohammad Malli Arab Open University, Beirut, Lebanon ACEEE Int. J. on Information Tehnology, Vol. 3, No. 2, June 2013 Fast Distribution of Repliated Content to Multi- Homed Clients Mohammad Malli Arab Open University, Beirut, Lebanon Email: mmalli@aou.edu.lb

More information

Learning Convention Propagation in BeerAdvocate Reviews from a etwork Perspective. Abstract

Learning Convention Propagation in BeerAdvocate Reviews from a etwork Perspective. Abstract CS 9 Projet Final Report: Learning Convention Propagation in BeerAdvoate Reviews from a etwork Perspetive Abstrat We look at the way onventions propagate between reviews on the BeerAdvoate dataset, and

More information

Multi-Channel Wireless Networks: Capacity and Protocols

Multi-Channel Wireless Networks: Capacity and Protocols Multi-Channel Wireless Networks: Capaity and Protools Tehnial Report April 2005 Pradeep Kyasanur Dept. of Computer Siene, and Coordinated Siene Laboratory, University of Illinois at Urbana-Champaign Email:

More information

Facility Location: Distributed Approximation

Facility Location: Distributed Approximation Faility Loation: Distributed Approximation Thomas Mosibroda Roger Wattenhofer Distributed Computing Group PODC 2005 Where to plae ahes in the Internet? A distributed appliation that has to dynamially plae

More information

Accommodations of QoS DiffServ Over IP and MPLS Networks

Accommodations of QoS DiffServ Over IP and MPLS Networks Aommodations of QoS DiffServ Over IP and MPLS Networks Abdullah AlWehaibi, Anjali Agarwal, Mihael Kadoh and Ahmed ElHakeem Department of Eletrial and Computer Department de Genie Eletrique Engineering

More information

On Dynamic Server Provisioning in Multi-channel P2P Live Streaming

On Dynamic Server Provisioning in Multi-channel P2P Live Streaming On Dynami Server Provisioning in Multi-hannel P2P Live Streaming Chuan Wu Baohun Li Shuqiao Zhao Department of Computer Siene Department of Eletrial Multimedia Development Group The University of Hong

More information

Distributed Resource Allocation Strategies for Achieving Quality of Service in Server Clusters

Distributed Resource Allocation Strategies for Achieving Quality of Service in Server Clusters Proeedings of the 45th IEEE Conferene on Deision & Control Manhester Grand Hyatt Hotel an Diego, CA, UA, Deember 13-15, 2006 Distributed Resoure Alloation trategies for Ahieving Quality of ervie in erver

More information

Outline: Software Design

Outline: Software Design Outline: Software Design. Goals History of software design ideas Design priniples Design methods Life belt or leg iron? (Budgen) Copyright Nany Leveson, Sept. 1999 A Little History... At first, struggling

More information

Performance Benchmarks for an Interactive Video-on-Demand System

Performance Benchmarks for an Interactive Video-on-Demand System Performane Benhmarks for an Interative Video-on-Demand System. Guo,P.G.Taylor,E.W.M.Wong,S.Chan,M.Zukerman andk.s.tang ARC Speial Researh Centre for Ultra-Broadband Information Networks (CUBIN) Department

More information

Announcements. Lecture Caching Issues for Multi-core Processors. Shared Vs. Private Caches for Small-scale Multi-core

Announcements. Lecture Caching Issues for Multi-core Processors. Shared Vs. Private Caches for Small-scale Multi-core Announements Your fous should be on the lass projet now Leture 17: Cahing Issues for Multi-ore Proessors This week: status update and meeting A short presentation on: projet desription (problem, importane,

More information

Flow Demands Oriented Node Placement in Multi-Hop Wireless Networks

Flow Demands Oriented Node Placement in Multi-Hop Wireless Networks Flow Demands Oriented Node Plaement in Multi-Hop Wireless Networks Zimu Yuan Institute of Computing Tehnology, CAS, China {zimu.yuan}@gmail.om arxiv:153.8396v1 [s.ni] 29 Mar 215 Abstrat In multi-hop wireless

More information

Multi-Piece Mold Design Based on Linear Mixed-Integer Program Toward Guaranteed Optimality

Multi-Piece Mold Design Based on Linear Mixed-Integer Program Toward Guaranteed Optimality INTERNATIONAL CONFERENCE ON MANUFACTURING AUTOMATION (ICMA200) Multi-Piee Mold Design Based on Linear Mixed-Integer Program Toward Guaranteed Optimality Stephen Stoyan, Yong Chen* Epstein Department of

More information

Automatic Physical Design Tuning: Workload as a Sequence Sanjay Agrawal Microsoft Research One Microsoft Way Redmond, WA, USA +1-(425)

Automatic Physical Design Tuning: Workload as a Sequence Sanjay Agrawal Microsoft Research One Microsoft Way Redmond, WA, USA +1-(425) Automati Physial Design Tuning: Workload as a Sequene Sanjay Agrawal Mirosoft Researh One Mirosoft Way Redmond, WA, USA +1-(425) 75-357 sagrawal@mirosoft.om Eri Chu * Computer Sienes Department University

More information

Drawing lines. Naïve line drawing algorithm. drawpixel(x, round(y)); double dy = y1 - y0; double dx = x1 - x0; double m = dy / dx; double y = y0;

Drawing lines. Naïve line drawing algorithm. drawpixel(x, round(y)); double dy = y1 - y0; double dx = x1 - x0; double m = dy / dx; double y = y0; Naïve line drawing algorithm // Connet to grid points(x0,y0) and // (x1,y1) by a line. void drawline(int x0, int y0, int x1, int y1) { int x; double dy = y1 - y0; double dx = x1 - x0; double m = dy / dx;

More information

What are Cycle-Stealing Systems Good For? A Detailed Performance Model Case Study

What are Cycle-Stealing Systems Good For? A Detailed Performance Model Case Study What are Cyle-Stealing Systems Good For? A Detailed Performane Model Case Study Wayne Kelly and Jiro Sumitomo Queensland University of Tehnology, Australia {w.kelly, j.sumitomo}@qut.edu.au Abstrat The

More information

On - Line Path Delay Fault Testing of Omega MINs M. Bellos 1, E. Kalligeros 1, D. Nikolos 1,2 & H. T. Vergos 1,2

On - Line Path Delay Fault Testing of Omega MINs M. Bellos 1, E. Kalligeros 1, D. Nikolos 1,2 & H. T. Vergos 1,2 On - Line Path Delay Fault Testing of Omega MINs M. Bellos, E. Kalligeros, D. Nikolos,2 & H. T. Vergos,2 Dept. of Computer Engineering and Informatis 2 Computer Tehnology Institute University of Patras,

More information

HEXA: Compact Data Structures for Faster Packet Processing

HEXA: Compact Data Structures for Faster Packet Processing Washington University in St. Louis Washington University Open Sholarship All Computer Siene and Engineering Researh Computer Siene and Engineering Report Number: 27-26 27 HEXA: Compat Data Strutures for

More information

Using Game Theory and Bayesian Networks to Optimize Cooperation in Ad Hoc Wireless Networks

Using Game Theory and Bayesian Networks to Optimize Cooperation in Ad Hoc Wireless Networks Using Game Theory and Bayesian Networks to Optimize Cooperation in Ad Ho Wireless Networks Giorgio Quer, Federio Librino, Lua Canzian, Leonardo Badia, Mihele Zorzi, University of California San Diego La

More information

SVC-DASH-M: Scalable Video Coding Dynamic Adaptive Streaming Over HTTP Using Multiple Connections

SVC-DASH-M: Scalable Video Coding Dynamic Adaptive Streaming Over HTTP Using Multiple Connections SVC-DASH-M: Salable Video Coding Dynami Adaptive Streaming Over HTTP Using Multiple Connetions Samar Ibrahim, Ahmed H. Zahran and Mahmoud H. Ismail Department of Eletronis and Eletrial Communiations, Faulty

More information

Pipelined Multipliers for Reconfigurable Hardware

Pipelined Multipliers for Reconfigurable Hardware Pipelined Multipliers for Reonfigurable Hardware Mithell J. Myjak and José G. Delgado-Frias Shool of Eletrial Engineering and Computer Siene, Washington State University Pullman, WA 99164-2752 USA {mmyjak,

More information

Extracting Partition Statistics from Semistructured Data

Extracting Partition Statistics from Semistructured Data Extrating Partition Statistis from Semistrutured Data John N. Wilson Rihard Gourlay Robert Japp Mathias Neumüller Department of Computer and Information Sienes University of Strathlyde, Glasgow, UK {jnw,rsg,rpj,mathias}@is.strath.a.uk

More information

NONLINEAR BACK PROJECTION FOR TOMOGRAPHIC IMAGE RECONSTRUCTION. Ken Sauer and Charles A. Bouman

NONLINEAR BACK PROJECTION FOR TOMOGRAPHIC IMAGE RECONSTRUCTION. Ken Sauer and Charles A. Bouman NONLINEAR BACK PROJECTION FOR TOMOGRAPHIC IMAGE RECONSTRUCTION Ken Sauer and Charles A. Bouman Department of Eletrial Engineering, University of Notre Dame Notre Dame, IN 46556, (219) 631-6999 Shool of

More information

Implementing Load-Balanced Switches With Fat-Tree Networks

Implementing Load-Balanced Switches With Fat-Tree Networks Implementing Load-Balaned Swithes With Fat-Tree Networks Hung-Shih Chueh, Ching-Min Lien, Cheng-Shang Chang, Jay Cheng, and Duan-Shin Lee Department of Eletrial Engineering & Institute of Communiations

More information

Design Implications for Enterprise Storage Systems via Multi-Dimensional Trace Analysis

Design Implications for Enterprise Storage Systems via Multi-Dimensional Trace Analysis Design Impliations for Enterprise Storage Systems via Multi-Dimensional Trae Analysis Yanpei Chen, Kiran Srinivasan, Garth Goodson, Randy Katz University of California, Berkeley, NetApp In. {yhen2, randy}@ees.berkeley.edu,

More information

Capturing Large Intra-class Variations of Biometric Data by Template Co-updating

Capturing Large Intra-class Variations of Biometric Data by Template Co-updating Capturing Large Intra-lass Variations of Biometri Data by Template Co-updating Ajita Rattani University of Cagliari Piazza d'armi, Cagliari, Italy ajita.rattani@diee.unia.it Gian Lua Marialis University

More information

Taming Decentralized POMDPs: Towards Efficient Policy Computation for Multiagent Settings

Taming Decentralized POMDPs: Towards Efficient Policy Computation for Multiagent Settings Taming Deentralized PMDPs: Towards ffiient Poliy omputation for Multiagent Settings. Nair and M. Tambe omputer Siene Dept. University of Southern alifornia Los Angeles A 90089 nair,tambe @us.edu M. Yokoo

More information

Acoustic Links. Maximizing Channel Utilization for Underwater

Acoustic Links. Maximizing Channel Utilization for Underwater Maximizing Channel Utilization for Underwater Aousti Links Albert F Hairris III Davide G. B. Meneghetti Adihele Zorzi Department of Information Engineering University of Padova, Italy Email: {harris,davide.meneghetti,zorzi}@dei.unipd.it

More information

The Happy Ending Problem

The Happy Ending Problem The Happy Ending Problem Neeldhara Misra STATUTORY WARNING This doument is a draft version 1 Introdution The Happy Ending problem first manifested itself on a typial wintery evening in 1933 These evenings

More information

User-level Fairness Delivered: Network Resource Allocation for Adaptive Video Streaming

User-level Fairness Delivered: Network Resource Allocation for Adaptive Video Streaming User-level Fairness Delivered: Network Resoure Alloation for Adaptive Video Streaming Mu Mu, Steven Simpson, Arsham Farshad, Qiang Ni, Niholas Rae Shool of Computing and Communiations, Lanaster University

More information

Semi-Supervised Affinity Propagation with Instance-Level Constraints

Semi-Supervised Affinity Propagation with Instance-Level Constraints Semi-Supervised Affinity Propagation with Instane-Level Constraints Inmar E. Givoni, Brendan J. Frey Probabilisti and Statistial Inferene Group University of Toronto 10 King s College Road, Toronto, Ontario,

More information

IN structured P2P overlay networks, each node and file key

IN structured P2P overlay networks, each node and file key 242 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 21, NO. 2, FEBRUARY 2010 Elasti Routing Table with Provable Performane for Congestion Control in DHT Networks Haiying Shen, Member, IEEE,

More information

Cluster Centric Fuzzy Modeling

Cluster Centric Fuzzy Modeling 10.1109/TFUZZ.014.300134, IEEE Transations on Fuzzy Systems TFS-013-0379.R1 1 Cluster Centri Fuzzy Modeling Witold Pedryz, Fellow, IEEE, and Hesam Izakian, Student Member, IEEE Abstrat In this study, we

More information

Improved Vehicle Classification in Long Traffic Video by Cooperating Tracker and Classifier Modules

Improved Vehicle Classification in Long Traffic Video by Cooperating Tracker and Classifier Modules Improved Vehile Classifiation in Long Traffi Video by Cooperating Traker and Classifier Modules Brendan Morris and Mohan Trivedi University of California, San Diego San Diego, CA 92093 {b1morris, trivedi}@usd.edu

More information

Graph-Based vs Depth-Based Data Representation for Multiview Images

Graph-Based vs Depth-Based Data Representation for Multiview Images Graph-Based vs Depth-Based Data Representation for Multiview Images Thomas Maugey, Antonio Ortega, Pasal Frossard Signal Proessing Laboratory (LTS), Eole Polytehnique Fédérale de Lausanne (EPFL) Email:

More information

arxiv: v1 [cs.db] 13 Sep 2017

arxiv: v1 [cs.db] 13 Sep 2017 An effiient lustering algorithm from the measure of loal Gaussian distribution Yuan-Yen Tai (Dated: May 27, 2018) In this paper, I will introdue a fast and novel lustering algorithm based on Gaussian distribution

More information

COST PERFORMANCE ASPECTS OF CCD FAST AUXILIARY MEMORY

COST PERFORMANCE ASPECTS OF CCD FAST AUXILIARY MEMORY COST PERFORMANCE ASPECTS OF CCD FAST AUXILIARY MEMORY Dileep P, Bhondarkor Texas Instruments Inorporated Dallas, Texas ABSTRACT Charge oupled devies (CCD's) hove been mentioned as potential fast auxiliary

More information

Gray Codes for Reflectable Languages

Gray Codes for Reflectable Languages Gray Codes for Refletable Languages Yue Li Joe Sawada Marh 8, 2008 Abstrat We lassify a type of language alled a refletable language. We then develop a generi algorithm that an be used to list all strings

More information

Adapting K-Medians to Generate Normalized Cluster Centers

Adapting K-Medians to Generate Normalized Cluster Centers Adapting -Medians to Generate Normalized Cluster Centers Benamin J. Anderson, Deborah S. Gross, David R. Musiant Anna M. Ritz, Thomas G. Smith, Leah E. Steinberg Carleton College andersbe@gmail.om, {dgross,

More information

Cluster-based Cooperative Communication with Network Coding in Wireless Networks

Cluster-based Cooperative Communication with Network Coding in Wireless Networks Cluster-based Cooperative Communiation with Network Coding in Wireless Networks Zygmunt J. Haas Shool of Eletrial and Computer Engineering Cornell University Ithaa, NY 4850, U.S.A. Email: haas@ee.ornell.edu

More information

We don t need no generation - a practical approach to sliding window RLNC

We don t need no generation - a practical approach to sliding window RLNC We don t need no generation - a pratial approah to sliding window RLNC Simon Wunderlih, Frank Gabriel, Sreekrishna Pandi, Frank H.P. Fitzek Deutshe Telekom Chair of Communiation Networks, TU Dresden, Dresden,

More information

Path Diversity for Overlay Multicast Streaming

Path Diversity for Overlay Multicast Streaming Path Diversity for Overlay Multiast Streaming Matulya Bansal and Avideh Zakhor Department of Eletrial Engineering and Computer Siene University of California, Berkeley Berkeley, CA 9472 {matulya, avz}@ees.berkeley.edu

More information

A DYNAMIC ACCESS CONTROL WITH BINARY KEY-PAIR

A DYNAMIC ACCESS CONTROL WITH BINARY KEY-PAIR Malaysian Journal of Computer Siene, Vol 10 No 1, June 1997, pp 36-41 A DYNAMIC ACCESS CONTROL WITH BINARY KEY-PAIR Md Rafiqul Islam, Harihodin Selamat and Mohd Noor Md Sap Faulty of Computer Siene and

More information

A Load-Balanced Clustering Protocol for Hierarchical Wireless Sensor Networks

A Load-Balanced Clustering Protocol for Hierarchical Wireless Sensor Networks International Journal of Advanes in Computer Networks and Its Seurity IJCNS A Load-Balaned Clustering Protool for Hierarhial Wireless Sensor Networks Mehdi Tarhani, Yousef S. Kavian, Saman Siavoshi, Ali

More information

Displacement-based Route Update Strategies for Proactive Routing Protocols in Mobile Ad Hoc Networks

Displacement-based Route Update Strategies for Proactive Routing Protocols in Mobile Ad Hoc Networks Displaement-based Route Update Strategies for Proative Routing Protools in Mobile Ad Ho Networks Mehran Abolhasan 1 and Tadeusz Wysoki 1 1 University of Wollongong, NSW 2522, Australia E-mail: mehran@titr.uow.edu.au,

More information

The Implementation of RRTs for a Remote-Controlled Mobile Robot

The Implementation of RRTs for a Remote-Controlled Mobile Robot ICCAS5 June -5, KINEX, Gyeonggi-Do, Korea he Implementation of RRs for a Remote-Controlled Mobile Robot Chi-Won Roh*, Woo-Sub Lee **, Sung-Chul Kang *** and Kwang-Won Lee **** * Intelligent Robotis Researh

More information

Uplink Channel Allocation Scheme and QoS Management Mechanism for Cognitive Cellular- Femtocell Networks

Uplink Channel Allocation Scheme and QoS Management Mechanism for Cognitive Cellular- Femtocell Networks 62 Uplink Channel Alloation Sheme and QoS Management Mehanism for Cognitive Cellular- Femtoell Networks Kien Du Nguyen 1, Hoang Nam Nguyen 1, Hiroaki Morino 2 and Iwao Sasase 3 1 University of Engineering

More information

A Multiagent Solution to Overcome Selfish Routing In Transportation Networks

A Multiagent Solution to Overcome Selfish Routing In Transportation Networks A Multiagent olution to Overome elfish Routing In Transportation Networks Mohammad Rashedul Hasan University of Nebraska-Linoln Linoln, NE 68588, UA Email: hasan@unl.edu Ana L. C. Bazzan PPGC / UFRG CP

More information

Colouring contact graphs of squares and rectilinear polygons de Berg, M.T.; Markovic, A.; Woeginger, G.

Colouring contact graphs of squares and rectilinear polygons de Berg, M.T.; Markovic, A.; Woeginger, G. Colouring ontat graphs of squares and retilinear polygons de Berg, M.T.; Markovi, A.; Woeginger, G. Published in: nd European Workshop on Computational Geometry (EuroCG 06), 0 Marh - April, Lugano, Switzerland

More information

Unsupervised Stereoscopic Video Object Segmentation Based on Active Contours and Retrainable Neural Networks

Unsupervised Stereoscopic Video Object Segmentation Based on Active Contours and Retrainable Neural Networks Unsupervised Stereosopi Video Objet Segmentation Based on Ative Contours and Retrainable Neural Networks KLIMIS NTALIANIS, ANASTASIOS DOULAMIS, and NIKOLAOS DOULAMIS National Tehnial University of Athens

More information

CleanUp: Improving Quadrilateral Finite Element Meshes

CleanUp: Improving Quadrilateral Finite Element Meshes CleanUp: Improving Quadrilateral Finite Element Meshes Paul Kinney MD-10 ECC P.O. Box 203 Ford Motor Company Dearborn, MI. 8121 (313) 28-1228 pkinney@ford.om Abstrat: Unless an all quadrilateral (quad)

More information

Cluster-Based Cumulative Ensembles

Cluster-Based Cumulative Ensembles Cluster-Based Cumulative Ensembles Hanan G. Ayad and Mohamed S. Kamel Pattern Analysis and Mahine Intelligene Lab, Eletrial and Computer Engineering, University of Waterloo, Waterloo, Ontario N2L 3G1,

More information

A Partial Sorting Algorithm in Multi-Hop Wireless Sensor Networks

A Partial Sorting Algorithm in Multi-Hop Wireless Sensor Networks A Partial Sorting Algorithm in Multi-Hop Wireless Sensor Networks Abouberine Ould Cheikhna Department of Computer Siene University of Piardie Jules Verne 80039 Amiens Frane Ould.heikhna.abouberine @u-piardie.fr

More information

KERNEL SPARSE REPRESENTATION WITH LOCAL PATTERNS FOR FACE RECOGNITION

KERNEL SPARSE REPRESENTATION WITH LOCAL PATTERNS FOR FACE RECOGNITION KERNEL SPARSE REPRESENTATION WITH LOCAL PATTERNS FOR FACE RECOGNITION Cuiui Kang 1, Shengai Liao, Shiming Xiang 1, Chunhong Pan 1 1 National Laboratory of Pattern Reognition, Institute of Automation, Chinese

More information

Algorithms for External Memory Lecture 6 Graph Algorithms - Weighted List Ranking

Algorithms for External Memory Lecture 6 Graph Algorithms - Weighted List Ranking Algorithms for External Memory Leture 6 Graph Algorithms - Weighted List Ranking Leturer: Nodari Sithinava Sribe: Andi Hellmund, Simon Ohsenreither 1 Introdution & Motivation After talking about I/O-effiient

More information

Gradient based progressive probabilistic Hough transform

Gradient based progressive probabilistic Hough transform Gradient based progressive probabilisti Hough transform C.Galambos, J.Kittler and J.Matas Abstrat: The authors look at the benefits of exploiting gradient information to enhane the progressive probabilisti

More information

Detecting Outliers in High-Dimensional Datasets with Mixed Attributes

Detecting Outliers in High-Dimensional Datasets with Mixed Attributes Deteting Outliers in High-Dimensional Datasets with Mixed Attributes A. Koufakou, M. Georgiopoulos, and G.C. Anagnostopoulos 2 Shool of EECS, University of Central Florida, Orlando, FL, USA 2 Dept. of

More information

Using Augmented Measurements to Improve the Convergence of ICP

Using Augmented Measurements to Improve the Convergence of ICP Using Augmented Measurements to Improve the onvergene of IP Jaopo Serafin, Giorgio Grisetti Dept. of omputer, ontrol and Management Engineering, Sapienza University of Rome, Via Ariosto 25, I-0085, Rome,

More information

Sparse Certificates for 2-Connectivity in Directed Graphs

Sparse Certificates for 2-Connectivity in Directed Graphs Sparse Certifiates for 2-Connetivity in Direted Graphs Loukas Georgiadis Giuseppe F. Italiano Aikaterini Karanasiou Charis Papadopoulos Nikos Parotsidis Abstrat Motivated by the emergene of large-sale

More information

Parametric Abstract Domains for Shape Analysis

Parametric Abstract Domains for Shape Analysis Parametri Abstrat Domains for Shape Analysis Xavier RIVAL (INRIA & Éole Normale Supérieure) Joint work with Bor-Yuh Evan CHANG (University of Maryland U University of Colorado) and George NECULA (University

More information

B4 and After: Managing Hierarchy, Partitioning, and Asymmetry for Availability and Scale in Google s Software-Defined WAN

B4 and After: Managing Hierarchy, Partitioning, and Asymmetry for Availability and Scale in Google s Software-Defined WAN B4 and After: Managing Hierarhy, Partitioning, and Asymmetry for Availability and Sale in Google s Software-Defined WAN Chi-Yao Hong Subhasree Mandal Mohammad Al-Fares Min Zhu Rihard Alimi Kondapa Naidu

More information

Analysis of input and output configurations for use in four-valued CCD programmable logic arrays

Analysis of input and output configurations for use in four-valued CCD programmable logic arrays nalysis of input and output onfigurations for use in four-valued D programmable logi arrays J.T. utler H.G. Kerkhoff ndexing terms: Logi, iruit theory and design, harge-oupled devies bstrat: s in binary,

More information

Detection and Recognition of Non-Occluded Objects using Signature Map

Detection and Recognition of Non-Occluded Objects using Signature Map 6th WSEAS International Conferene on CIRCUITS, SYSTEMS, ELECTRONICS,CONTROL & SIGNAL PROCESSING, Cairo, Egypt, De 9-31, 007 65 Detetion and Reognition of Non-Oluded Objets using Signature Map Sangbum Park,

More information

13.1 Numerical Evaluation of Integrals Over One Dimension

13.1 Numerical Evaluation of Integrals Over One Dimension 13.1 Numerial Evaluation of Integrals Over One Dimension A. Purpose This olletion of subprograms estimates the value of the integral b a f(x) dx where the integrand f(x) and the limits a and b are supplied

More information

Discrete sequential models and CRFs. 1 Case Study: Supervised Part-of-Speech Tagging

Discrete sequential models and CRFs. 1 Case Study: Supervised Part-of-Speech Tagging 0-708: Probabilisti Graphial Models 0-708, Spring 204 Disrete sequential models and CRFs Leturer: Eri P. Xing Sribes: Pankesh Bamotra, Xuanhong Li Case Study: Supervised Part-of-Speeh Tagging The supervised

More information

An Efficient and Scalable Approach to CNN Queries in a Road Network

An Efficient and Scalable Approach to CNN Queries in a Road Network An Effiient and Salable Approah to CNN Queries in a Road Network Hyung-Ju Cho Chin-Wan Chung Dept. of Eletrial Engineering & Computer Siene Korea Advaned Institute of Siene and Tehnology 373- Kusong-dong,

More information

RAC 2 E: Novel Rendezvous Protocol for Asynchronous Cognitive Radios in Cooperative Environments

RAC 2 E: Novel Rendezvous Protocol for Asynchronous Cognitive Radios in Cooperative Environments 21st Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communiations 1 RAC 2 E: Novel Rendezvous Protool for Asynhronous Cognitive Radios in Cooperative Environments Valentina Pavlovska,

More information

Methods for Multi-Dimensional Robustness Optimization in Complex Embedded Systems

Methods for Multi-Dimensional Robustness Optimization in Complex Embedded Systems Methods for Multi-Dimensional Robustness Optimization in Complex Embedded Systems Arne Hamann, Razvan Rau, Rolf Ernst Institute of Computer and Communiation Network Engineering Tehnial University of Braunshweig,

More information

Constructing Transaction Serialization Order for Incremental. Data Warehouse Refresh. Ming-Ling Lo and Hui-I Hsiao. IBM T. J. Watson Research Center

Constructing Transaction Serialization Order for Incremental. Data Warehouse Refresh. Ming-Ling Lo and Hui-I Hsiao. IBM T. J. Watson Research Center Construting Transation Serialization Order for Inremental Data Warehouse Refresh Ming-Ling Lo and Hui-I Hsiao IBM T. J. Watson Researh Center July 11, 1997 Abstrat In typial pratie of data warehouse, the

More information

the data. Structured Principal Component Analysis (SPCA)

the data. Structured Principal Component Analysis (SPCA) Strutured Prinipal Component Analysis Kristin M. Branson and Sameer Agarwal Department of Computer Siene and Engineering University of California, San Diego La Jolla, CA 9193-114 Abstrat Many tasks involving

More information

Cooperative Coverage Extension for Relay-Union Networks

Cooperative Coverage Extension for Relay-Union Networks 1.119/TPDS.214.23821, IEEE Transations on Parallel and Distributed Systems IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS 1 Cooperative Coverage Extension for Relay-Union Networks Yong Cui, Xiao

More information

THROUGHPUT EVALUATION OF AN ASYMMETRICAL FDDI TOKEN RING NETWORK WITH MULTIPLE CLASSES OF TRAFFIC

THROUGHPUT EVALUATION OF AN ASYMMETRICAL FDDI TOKEN RING NETWORK WITH MULTIPLE CLASSES OF TRAFFIC THROUGHPUT EVALUATION OF AN ASYMMETRICAL FDDI TOKEN RING NETWORK WITH MULTIPLE CLASSES OF TRAFFIC Priya N. Werahera and Anura P. Jayasumana Department of Eletrial Engineering Colorado State University

More information

C 2 C 3 C 1 M S. f e. e f (3,0) (0,1) (2,0) (-1,1) (1,0) (-1,0) (1,-1) (0,-1) (-2,0) (-3,0) (0,-2)

C 2 C 3 C 1 M S. f e. e f (3,0) (0,1) (2,0) (-1,1) (1,0) (-1,0) (1,-1) (0,-1) (-2,0) (-3,0) (0,-2) SPECIAL ISSUE OF IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION: MULTI-ROBOT SSTEMS, 00 Distributed reonfiguration of hexagonal metamorphi robots Jennifer E. Walter, Jennifer L. Welh, and Nany M. Amato Abstrat

More information

The Minimum Redundancy Maximum Relevance Approach to Building Sparse Support Vector Machines

The Minimum Redundancy Maximum Relevance Approach to Building Sparse Support Vector Machines The Minimum Redundany Maximum Relevane Approah to Building Sparse Support Vetor Mahines Xiaoxing Yang, Ke Tang, and Xin Yao, Nature Inspired Computation and Appliations Laboratory (NICAL), Shool of Computer

More information

Self-Adaptive Parent to Mean-Centric Recombination for Real-Parameter Optimization

Self-Adaptive Parent to Mean-Centric Recombination for Real-Parameter Optimization Self-Adaptive Parent to Mean-Centri Reombination for Real-Parameter Optimization Kalyanmoy Deb and Himanshu Jain Department of Mehanial Engineering Indian Institute of Tehnology Kanpur Kanpur, PIN 86 {deb,hjain}@iitk.a.in

More information

Dynamic Backlight Adaptation for Low Power Handheld Devices 1

Dynamic Backlight Adaptation for Low Power Handheld Devices 1 Dynami Baklight Adaptation for ow Power Handheld Devies 1 Sudeep Pasriha, Manev uthra, Shivajit Mohapatra, Nikil Dutt and Nalini Venkatasubramanian 444, Computer Siene Building, Shool of Information &

More information

特集 Road Border Recognition Using FIR Images and LIDAR Signal Processing

特集 Road Border Recognition Using FIR Images and LIDAR Signal Processing デンソーテクニカルレビュー Vol. 15 2010 特集 Road Border Reognition Using FIR Images and LIDAR Signal Proessing 高木聖和 バーゼル ファルディ Kiyokazu TAKAGI Basel Fardi ヘンドリック ヴァイゲル Hendrik Weigel ゲルド ヴァニーリック Gerd Wanielik This paper

More information

Rotation Invariant Spherical Harmonic Representation of 3D Shape Descriptors

Rotation Invariant Spherical Harmonic Representation of 3D Shape Descriptors Eurographis Symposium on Geometry Proessing (003) L. Kobbelt, P. Shröder, H. Hoppe (Editors) Rotation Invariant Spherial Harmoni Representation of 3D Shape Desriptors Mihael Kazhdan, Thomas Funkhouser,

More information

Alleviating DFT cost using testability driven HLS

Alleviating DFT cost using testability driven HLS Alleviating DFT ost using testability driven HLS M.L.Flottes, R.Pires, B.Rouzeyre Laboratoire d Informatique, de Robotique et de Miroéletronique de Montpellier, U.M. CNRS 5506 6 rue Ada, 34392 Montpellier

More information

DoS-Resistant Broadcast Authentication Protocol with Low End-to-end Delay

DoS-Resistant Broadcast Authentication Protocol with Low End-to-end Delay DoS-Resistant Broadast Authentiation Protool with Low End-to-end Delay Ying Huang, Wenbo He and Klara Nahrstedt {huang, wenbohe, klara}@s.uiu.edu Department of Computer Siene University of Illinois at

More information

Recommendation Subgraphs for Web Discovery

Recommendation Subgraphs for Web Discovery Reommation Subgraphs for Web Disovery Arda Antikaioglu Department of Mathematis Carnegie Mellon University aantika@andrew.mu.edu R. Ravi Tepper Shool of Business Carnegie Mellon University ravi@mu.edu

More information

A Comparison of Hard-state and Soft-state Signaling Protocols

A Comparison of Hard-state and Soft-state Signaling Protocols University of Massahusetts Amherst SholarWorks@UMass Amherst Computer Siene Department Faulty Publiation Series Computer Siene 2003 A Comparison of Hard-state and Soft-state Signaling Protools Ping Ji

More information

Algorithms, Mechanisms and Procedures for the Computer-aided Project Generation System

Algorithms, Mechanisms and Procedures for the Computer-aided Project Generation System Algorithms, Mehanisms and Proedures for the Computer-aided Projet Generation System Anton O. Butko 1*, Aleksandr P. Briukhovetskii 2, Dmitry E. Grigoriev 2# and Konstantin S. Kalashnikov 3 1 Department

More information

Exploiting Enriched Contextual Information for Mobile App Classification

Exploiting Enriched Contextual Information for Mobile App Classification Exploiting Enrihed Contextual Information for Mobile App Classifiation Hengshu Zhu 1 Huanhuan Cao 2 Enhong Chen 1 Hui Xiong 3 Jilei Tian 2 1 University of Siene and Tehnology of China 2 Nokia Researh Center

More information

Plot-to-track correlation in A-SMGCS using the target images from a Surface Movement Radar

Plot-to-track correlation in A-SMGCS using the target images from a Surface Movement Radar Plot-to-trak orrelation in A-SMGCS using the target images from a Surfae Movement Radar G. Golino Radar & ehnology Division AMS, Italy ggolino@amsjv.it Abstrat he main topi of this paper is the formulation

More information

A Dual-Hamiltonian-Path-Based Multicasting Strategy for Wormhole-Routed Star Graph Interconnection Networks

A Dual-Hamiltonian-Path-Based Multicasting Strategy for Wormhole-Routed Star Graph Interconnection Networks A Dual-Hamiltonian-Path-Based Multiasting Strategy for Wormhole-Routed Star Graph Interonnetion Networks Nen-Chung Wang Department of Information and Communiation Engineering Chaoyang University of Tehnology,

More information

Cross-layer Resource Allocation on Broadband Power Line Based on Novel QoS-priority Scheduling Function in MAC Layer

Cross-layer Resource Allocation on Broadband Power Line Based on Novel QoS-priority Scheduling Function in MAC Layer Communiations and Networ, 2013, 5, 69-73 http://dx.doi.org/10.4236/n.2013.53b2014 Published Online September 2013 (http://www.sirp.org/journal/n) Cross-layer Resoure Alloation on Broadband Power Line Based

More information

Sequential Incremental-Value Auctions

Sequential Incremental-Value Auctions Sequential Inremental-Value Autions Xiaoming Zheng and Sven Koenig Department of Computer Siene University of Southern California Los Angeles, CA 90089-0781 {xiaominz,skoenig}@us.edu Abstrat We study the

More information

Scalable TCP: Improving Performance in Highspeed Wide Area Networks

Scalable TCP: Improving Performance in Highspeed Wide Area Networks Salable TP: Improving Performane in Highspeed Wide Area Networks Tom Kelly ERN - IT Division Geneva 3 Switzerland tk@am.a.uk ABSTRAT TP ongestion ontrol an perform badly in highspeed wide area networks

More information

PBFT: A Byzantine Renaissance. The Setup. What could possibly go wrong? The General Idea. Practical Byzantine Fault-Tolerance (CL99, CL00)

PBFT: A Byzantine Renaissance. The Setup. What could possibly go wrong? The General Idea. Practical Byzantine Fault-Tolerance (CL99, CL00) PBFT: A Byzantine Renaissane Pratial Byzantine Fault-Tolerane (CL99, CL00) first to be safe in asynhronous systems live under weak synhrony assumptions -Byzantine Paos! The Setup Crypto System Model Asynhronous

More information

Approximate logic synthesis for error tolerant applications

Approximate logic synthesis for error tolerant applications Approximate logi synthesis for error tolerant appliations Doohul Shin and Sandeep K. Gupta Eletrial Engineering Department, University of Southern California, Los Angeles, CA 989 {doohuls, sandeep}@us.edu

More information

A Novel Validity Index for Determination of the Optimal Number of Clusters

A Novel Validity Index for Determination of the Optimal Number of Clusters IEICE TRANS. INF. & SYST., VOL.E84 D, NO.2 FEBRUARY 2001 281 LETTER A Novel Validity Index for Determination of the Optimal Number of Clusters Do-Jong KIM, Yong-Woon PARK, and Dong-Jo PARK, Nonmembers

More information

Smooth Trajectory Planning Along Bezier Curve for Mobile Robots with Velocity Constraints

Smooth Trajectory Planning Along Bezier Curve for Mobile Robots with Velocity Constraints Smooth Trajetory Planning Along Bezier Curve for Mobile Robots with Veloity Constraints Gil Jin Yang and Byoung Wook Choi Department of Eletrial and Information Engineering Seoul National University of

More information

A Unified Subdivision Scheme for Polygonal Modeling

A Unified Subdivision Scheme for Polygonal Modeling EUROGRAPHICS 2 / A. Chalmers and T.-M. Rhyne (Guest Editors) Volume 2 (2), Number 3 A Unified Subdivision Sheme for Polygonal Modeling Jérôme Maillot Jos Stam Alias Wavefront Alias Wavefront 2 King St.

More information

Abstract. Key Words: Image Filters, Fuzzy Filters, Order Statistics Filters, Rank Ordered Mean Filters, Channel Noise. 1.

Abstract. Key Words: Image Filters, Fuzzy Filters, Order Statistics Filters, Rank Ordered Mean Filters, Channel Noise. 1. Fuzzy Weighted Rank Ordered Mean (FWROM) Filters for Mixed Noise Suppression from Images S. Meher, G. Panda, B. Majhi 3, M.R. Meher 4,,4 Department of Eletronis and I.E., National Institute of Tehnology,

More information

Boosted Random Forest

Boosted Random Forest Boosted Random Forest Yohei Mishina, Masamitsu suhiya and Hironobu Fujiyoshi Department of Computer Siene, Chubu University, 1200 Matsumoto-ho, Kasugai, Aihi, Japan {mishi, mtdoll}@vision.s.hubu.a.jp,

More information

A Multi-Head Clustering Algorithm in Vehicular Ad Hoc Networks

A Multi-Head Clustering Algorithm in Vehicular Ad Hoc Networks International Journal of Computer Theory and Engineering, Vol. 5, No. 2, April 213 A Multi-Head Clustering Algorithm in Vehiular Ad Ho Networks Shou-Chih Lo, Yi-Jen Lin, and Jhih-Siao Gao Abstrat Clustering

More information

System-Level Parallelism and Throughput Optimization in Designing Reconfigurable Computing Applications

System-Level Parallelism and Throughput Optimization in Designing Reconfigurable Computing Applications System-Level Parallelism and hroughput Optimization in Designing Reonfigurable Computing Appliations Esam El-Araby 1, Mohamed aher 1, Kris Gaj 2, arek El-Ghazawi 1, David Caliga 3, and Nikitas Alexandridis

More information

Dynamic Algorithms Multiple Choice Test

Dynamic Algorithms Multiple Choice Test 3226 Dynami Algorithms Multiple Choie Test Sample test: only 8 questions 32 minutes (Real test has 30 questions 120 minutes) Årskort Name Eah of the following 8 questions has 4 possible answers of whih

More information

COSSIM An Integrated Solution to Address the Simulator Gap for Parallel Heterogeneous Systems

COSSIM An Integrated Solution to Address the Simulator Gap for Parallel Heterogeneous Systems COSSIM An Integrated Solution to Address the Simulator Gap for Parallel Heterogeneous Systems Andreas Brokalakis Synelixis Solutions Ltd, Greee brokalakis@synelixis.om Nikolaos Tampouratzis Teleommuniation

More information