RAPIER: Integrating Routing and Scheduling for Coflow-aware Data Center Networks

Size: px
Start display at page:

Download "RAPIER: Integrating Routing and Scheduling for Coflow-aware Data Center Networks"

Transcription

1 : Inegraing Rouing and cheduling for Coflow-aware Daa Cener Neworks Yangming Zhao (UETC), ai Chen, Wei Bai, Minlan Yu (UC), Chen Tian (HUT), Yanhui Geng (Huawei), Yiming Zhang (NUDT), Dan Li (Tsinghua), heng Wang (UETC) ING Group, HUT Absrac In he daa flow models of oday s daa cener applicaions such as MapReduce, park and Dryad, muliple flows can comprise a coflow group semanically. Only compleing all flows in a coflow is meaningful o an applicaion. To opimize applicaion performance, rouing and scheduling mus be oinly considered a he level of a coflow raher han individual flows. However, prior soluions have significan limiaion: hey only consider scheduling, which is insufficien. To his end, we presen, a coflow-aware nework opimizaion framework ha seamlessly inegraes rouing and scheduling for beer applicaion performance. Using a smallscale esbed implemenaion and large-scale simulaions, we demonsrae ha significanly reduces he average coflow compleion ime (CCT) by up o 79.% compared o he sae-of-he-ar scheduling-only soluion, and i is readily implemenable wih exising commodiy swiches. I. INTRODUCTION Cluser compuing frameworks such as MapReduce [], Dryad [8], park [] and so on have become he mainsream plaforms for daa processing and analysis in oday s cloud services. A common feaure of hese differen compuing paradigms is ha hey all implemen a daa flow compuing model, in which a group of daa flows need o pass hrough a sequence of inermediae processing sages before generaing he final resuls. These inermediae flow ransfers can accoun for more han % of ob compleion ime [8], and have a significan impac on ob performance. Therefore, opimizing such flow ransfers is imporan for applicaions. The erm coflow is defined as he se of all flows ransferring daa beween wo sages of a ob [7]. To opimize applicaion performance, we need o opimize flow ransfers a he level of coflow raher han individual ones. This is because he ob compleion ime depends on he ime i akes o complee he enire coflow, insead of he ime o complee individual flows composing i. For example, in MapReduce [] and BP [], a sage canno complee, or someimes even sar, before i receives all he flows in a coflow from he previous sage. From an applicaion s perspecive, when a sage is pending for he inpu daa, he CPU ofen sis idle or is under-uilized. As a resul, reducing he coflow compleion ime (CCT) can furher improve CPU uilizaion, maximizing applicaion performance and ob hroughpu in a given ime period. To minimize average CCT, boh rouing and scheduling mus be considered simulaneously (see ecion II for deails). This work was performed when Yangming Zhao was an inern suden a he ING Group of HUT under supervision of Prof. ai Chen. However, prior soluions for nework flow opimizaion such as [,, 8, 9,,,, ] have significan limiaions (Table I): some of hem (e.g., [,,,, ]) are ineffecive, because hey are coflow-agnosic ha do no accoun for collecive behaviors of flows belonging o a coflow; exising coflow-aware soluions (e.g., [8, 9, ]) are insufficien, because alhough hese approaches improve by saring o consider coflow level semanics, hey only focus on scheduling while neglecing an indispensable componen rouing. We show in ecion V ha his can direcly lead o.7% performance loss. Relaed work Coflow-aware Rouing cheduling pfabric [], PDQ [], Pase [], D [], ec. No No Yes Varys [9], Baraa [] Orchesra [8] Yes No Yes Yes Yes Yes TABLE I UMMARY OF PRIOR OLUTION AND COMPARION TO Moivaed by his siuaion, we design and implemen, a coflow-aware nework opimizaion sysem for daa cener neworks (DCNs). To improve average CCT, seamlessly combines rouing and scheduling ogeher by formulaing i as a oin opimizaion model. This model is a nonlinear programming and conains ineger variables; i is impossible o be direcly solved. Accordingly, we propose an efficien heurisic o approximaely solve his problem based on he relaxaion of he model. We evaluae using a small-scale esbed implemenaion as well as large-scale simulaions. Our evaluaion resuls show ha can reduce average CCT by up o 79.% and.%, compared o he scheduling-only and rouing-only schemes respecively. Our implemenaion verifies ha RAPI- ER can be readily implemenable wih exising commodiy swiches. In summary, he main highlighs of his paper include: A key observaion ha boh rouing and scheduling mus be oinly considered for opimizing average CCT. (ecion II) A coflow-aware nework opimizaion soluion,, which akes ino accoun rouing and scheduling simulaneously for he firs ime. In he course of sysem design, we also develop fas and efficien online algorihms o approximaely solve heoreical NP-hard problems.

2 M d u D f u D a f a Group Mb u Mcomplees u Group b complees d D d D =Mb =Mb f f a b f f a b f a =Mb f a =Mb f a =Mb f a =Mb f a =Mb Group a complees MGroup d M a complees d D D =M =M Group a complees Group a complees M u M u f a =Mb f a =Mb f a =Mb f a =Mb =Mb =Mb D D M d M d =M =M f a f a =Mb u D M u f a Group a complees M d =Mb =M d D f a M d D =M f a =Mb D f a =Mb =Mb M cheduling on cheduling f a on f a cheduling on cheduling f on M b (a) A possible case by ECMP cheduling (b) Rouing on M wihou u D coflow concep (c) OpimalCoflow rouingb complees u D Coflow a complees M u M f a =Mb u f a =Mb f a f a f u D cheduling on M u D u D cheduling cheduling on on M M u D u D b f a f a f =M f u D b =M f a a f a f a fgroup a b complees D Group b complees f a D =Mb d D Group b complees =Mb Coflow d D a complees f d D a cheduling on M d D d D cheduling on on M d D f a M d M d f a =Mb f a =Mb f d D f a b f f a a Group a complees f a Group a complees Coflow b complees Group a complees cheduling Coflow on a M complees Coflow b complees Coflow Coflow b complees complees Coflow a complees Coflow a complees u D u D Coflow b complees f (d) Opimal a scheduling up on (a) f a (e) Opimal scheduling up on (b) (f) Opimal scheduling up on (c) u D u D Fig.. Af a moivaingf example, where (a) (c) show differen rouing schemes, and (d) (f) show he opimal scheduling Group b complees u D schemes for (a) (c). a f b Group b complees d D f a (ecion cheduling III and on M ecion d D IV). A real esbed f f a b implemenaion f f and exensive large-scale cheduling M a b d D cheduling simulaions. on M d D (ecion V) f f d D a Group b a complees f a II. A MOTIVATING Group a complees f a EXAMPLE Coflow a complees In his secion, we make Coflow keyb complees observaions hrough a moivaing example in Fig. Coflow. b In complees his example, here are wo he average CCT. Coflow a complees Coflow b complees Coflow a complees coflows: Coflow a has u D flows f a and f a wih he sizes of u D f a f u D Mb and Mb respecively; b Coflow f u D b has flows a and f a u D f a f f a b wih he sizes of Mb and Mb respecively; and he f a f a link bandwidhs are all Mbps. As a reference poin, he Coflow a complees opimal average cheduling CCT on Mof d D his example should be.s. d D Coflow a complees The firscheduling akeaway on M d D f a is ha scheduling f d D b alone is no sufficien design below. d D o opimize average f a CCT. When he rouing is fixed, good III. DEIGN OVERVIEW f a f scheduling can minimize he b Coflow average b complees CCT by deermining Coflow b complees Coflow a complees he sequence of flows o sendcoflow oub raffic. complees Fig.(a) shows a Coflow b complees Coflow a complees Coflow a complees case of randomized rouing by equal-cos mulipah (ECMP). Coflow b complees Wih a naive cheduling scheduling on M u D such as fair sharing, boh coflows are dominaed by cheduling pah f on M a M u D d D hence heir CCT are boh s. If using he opimal f a scheduling shown in Fig.(d), he CCTs for wo coflows become s and s respecively; apparenly, schedulingcheduling does play on M a criical d D role. However, he average CCT (in his rouing) cheduling is only f M.s, d D which sill has a.s gap o he a real opimal value.s. I is clear ha rouing should also f a play a criical role: Coflow he loads a complees of wo pahs in Fig.(a) are severely unbalanced, where Coflow pah b complees M Coflow a complees d D has a raffic load doubles ha of pah Coflow M u b complees D. The second akeaway is ha considering rouing and scheduling separaely canno opimize average CCT. As an example in Fig.(b), a load-balancing rouing resuls in: boh flows of Coflow a are roued on M u D while he flows of Coflow b on M d D; now he nework is more balanced. However, he opimal CCTs for coflows a and b in his case are.s and.s respecively (see Fig.(e)); he average CCT.s is sill no opimal. The reason is ha flows of he same coflow are roued hrough he same pah, which leaves lile space for scheduling o ake effec for reducing The conclusion is ha boh rouing and scheduling mus be oinly considered in order o opimize average CCT. In our example, he minimal average CCT can be achieved by combining he rouing in Fig.(c) and scheduling in Fig.(f). In his case, he CCTs of wo coflows are s and.s respecively, and he average CCT is minimized. This moivaes our opimizes he average CCT in daa-inensive DCNs by coordinaing rouing and scheduling flows in he neworks. Given each coflow wih informaion abou is individual flows, such as flow sizes, and sources/desinaions, deermines which pahs o carry hese flows, when o sar hem, and a wha rae o serve hem, in order o opimize he average CCT of all he coflows in he neworks. Inspired by [8, 9], we design o work in a cenralized, cooperaive manner. This decision is also coheren wih many recen cenralized daa cener designs such as [,,,,,, ], ec. As prior works [, 9,,, ], assumes ha he informaion abou a coflow can be readily derived from upper layer applicaions [7] or using sae-of-he-ar predicion echniques []. A. Desirable Properies We idenify he following goals when designing. cheduling on f a cheduling on

3 Algorihm : The Framework : Procedure MinimizeMeanCCT(Coflows Ω, Bandwidh R) : or all he coflows in Ω non-increasingly according o heir waiing ime; Ω Ω : while Ω Φ do : T min, C min Φ; : for C Ω do : T C =MinimumCCT(C, R); /* compue he minimum compleion ime for coflow C, and he corresponding rouing and rae allocaion */ 7: if C.waiT ime() > δ hen 8: T min T C, C min C; 9: break; : end if : if T C < T min hen : T min T C, C min C; : end if : end for : Ω Ω \ C min ; : Assign all he flows in coflow C min using rouing and raes compued in Line, and hen updae R; 7: end while 8: DisribueBandwidh(Ω, R); /* disribue he remaining bandwidh for work conservaion */ 9: end procedure calabiliy: is necessarily an online sysem. Up on a new incoming coflow, he algorihms mus be able o quickly and efficienly decide he rouing pahs, raes, and scheduling orders for all individual flows in he coflow. For his purpose, hese algorihms mus run in real-ime wih low ime complexiy. arvaion-free: As allows bandwidh preempion, we mus ensure ha any coflow should no sarve for an arbirarily long period, hough his migh benefi he average CCT in he nework. Work-conserving: Work-conservaion means ha he nework resource sis idle only if here is no raffic demand in he nework. We require o be workconserving o fully uilize nework capaciy and o minimize CCT. Readily deployable: The sysem should be readily implemenable wih exising commodiy swiches and easy o deploy wihou modifying any nework devices. Ensure coexisence: The sysem mus be able o work wih all ypes of raffic. Especially, laency-sensiive ineracive raffic mus be delivered wihou any delay. B. in a Nushell A a high level, o achieve scalabiliy, mainly orchesraes large coflows of daa-inensive applicaions, while laency-sensiive individual flows and small coflows are reaed as background raffic; background raffic can be sen direcly and roued over he nework using ECMP. A sie broker periodically predics he usage of background raffic in each link, and derives he residual bandwidh for coflow scheduling. We describe he coflow opimizaion framework of wih Algorihm, which is invoked whenever a new coflow comes or an exising coflow finishes. More specifically, when a new coflow arrives, is riggered o compue he rouing and he ransmission rae for each individual flow (we allow bandwidh preempion as below). When an exising coflow finishes and nework resource is released, we also need o rigger o deermine which coflows should ake up he released bandwidh. The underlying scheduling policy assumes is he well-known minimum remaining ime firs (MRTF) [9, ]. As he inpu of Algorihm, all he coflows ha are no compleed should be included in Ω. In his case, even if a coflow is occupying he bandwidh in he nework, i may be preemped if a smaller coflow comes. On he oher hand, if par of a coflow is served, is remaining volume informaion should be updaed when we recompue he coflow order. To preven sarvaion, prioriizes coflows which are waiing for a ime longer han a user-defined hreshold o schedule (Line 7-). Oher han ha, i is he urn for he coflow wih he minimum compleion ime (Line -7). When a coflow is seleced o send, updaes he bandwidh uilizaion and coninues o find he nex coflow wih he nex minimum compleion ime (hrough Line -). Afer he schedule order is deermined, he remaining bandwidh is disribued o differen coflows for he work-conservaion purpose (Line 8). Noe ha here are wo key algorihms in. The firs one is o calculae he minimum compleion ime for each coflow given he informaion of all individual flows in his coflow and he nework resource ha can be used (Line ). The oher one is o disribue he remaining bandwidh for work-conservaion (Line 8). Designing hese wo complex algorihms is challenging. IV. ALGORITHM DETAIL In his secion, we presen he deails for he wo key algorihms in. In ecion IV-A, we discuss how o calculae he minimum compleion ime for a single coflow by oinly opimizing rouing and scheduling. Afer ha, we analyze he approximaion raio of our algorihm in ecion IV-B. In ecion IV-C, we presen he heurisic algorihm in o disribue he remaining bandwidh o flows for work-conservaion. A. Minimize ingle Coflow Compleion Time Given he informaion of all he flows in a coflow, such as flow volume, source, desinaion, and nework resource (he residual bandwidh on each link), we can formulae he problem o minimize he CCT of a coflow i as follows: minimize i ()

4 ubec o: v i = i, (a) b i b i x k i R l l (b) x k i =, (c) k x k i {, },, k (d) i in he obecive is he compleion ime of coflow i, and hence i should be minimized. ymbol v i is he flow volume of he h flow in coflow i, while b i is he bandwidh assigned o his flow. Wih consrain (a), we direcly enforce ha he compleion ime of all he flows o equal o he CCT, since i is reasonable o have all he flows in a coflow o have he same compleion ime (aka, he boleneck s compleion ime) in he opimum soluion [7] [9]. Le x k i indicae wheher he h flow in coflow i uses is k h pah (he link se of his pah is denoed by p k i), he lef-hand erm of (b) calculae he capaciy ha is used by coflow i on link l, which should be less han he residual capaciy of link l, denoed by R l. (c) and (d) require ha a flow only chooses one rouing pah. I is impossible o solve problem () direcly, since his programming no only is nonlinear, bu also has binary ineger variables. This problem is an ineger muli-commodiy flow problem ha is proven o be NP-hard []. Therefore, we resor o designing an efficien heurisic o solve his problem. Based on consrain (a), we know ha he rae of each flow b i is direcly proporional o is volume v i, i.e., b i = α i v i. The larger α i means more bandwidh is obained by flows in coflow i, and hence he smaller compleion ime is required. In fac, we have α i = / i. The programming model () can be modified as follows: ubec o: maximize α i () α i v i x k i R l l (a) (c), (d) However, here are sill binary variables x k i in programming model (), which leads he problem o be inracable on large scale sysems. Therefore, we furher relax he binary consrain and obain: maximize α i () ubec o: v i (α i x k i) R l l (a) x k i l (b) (c) I should be noed ha here is a produc of wo variables (i.e., α i and x k i) in consrain (a). I makes he problem difficul o solve since i is a concave opimizaion. To solve his problem, variables m k i are inroduced o subsiue his produc and we obain: ubec o: maximize α i () v i m k i R l l (a) m k i = α i (b) k m k i l (c) Now, problem () becomes a linear programming ha has only n i + variables and L + F i consrains (n i is he number of candidae pah for h flow in coflow i, L is he number of links, and F i is he number of flows in coflow i). This is a small scale linear programming and can be solved in a imely manner. However, since we relax he binary ineger consrain, he soluion may be ha some x k i are decimal facions. To solve his problem, we roue he h flow in coflow i o a pah k such ha m k i = max k m k i. When he pah of each flow is deermined, i.e., x k i is fixed, we can go back o problem (). We subsiue in he obained x k i values, and make () a linear programming problem and solve i. Given he pah of each flow, he minimum CCT is exacly he inverse of he obecive in (). In summary, he heurisic is: inegrae scheduling and rouing in he opimizaion ogeher and le scheduling guide he rouing selecion; afer fixing he rouing wih he approximaion, he opimal scheduling is hen derived. The heurisic o pursue he minimum compleion ime of a coflow is summarized as Algorihm. B. Approximaion Bound Analysis We have presened a heurisic o pursue he minimum CCT by relaxing problem (). In his par, we show how good he performance of he algorihm is hrough heoreical analysis. Algorihm : Minimize Coflow Compleion Time : Procedure MinimumCCT(Coflow C i, Bandwidh R) : olve problem () wih coflow and nework resource uilizaion informaion : for all flow in coflow C i do : Iniialize x k i for all k : k arg k max m k i : x k i 7: end for 8: olve () by fixing x k i o be he resul obained from Line -7 9: b i α i v i, where α i is he obecive in Line 8 : i α : reurn i : end procedure

5 Theorem : Assume he minimum CCT is min and alg is he CCT obained by Algorihm, hen alg min where is he number of candidae pahs for each flow. Proof: To prove his heorem, he equivalen proposiion is α alg αmax where α alg and α max are he inverse of alg and min, respecively. Assume he obecive of problem () is α upper, here mus be α upper α max () From Algorihm, we roue each flow o he pah wih maximum m k i, and hence we have m k i αupper xk i () for any k. ubsiue () ino he consrains of problem (), we have α upper v ix k i m i v i R l Combine wih he fac ha he consrain (c) and (d) are guaraneed by Algorihm, we know ha αupper is a feasible soluion o problem () for he he given x k i. I means ha α alg αupper αmax I is worh noing ha alhough he heoreical bound is loose, in pracice our implemenaion obains very good resuls. C. Disribue Bandwidh for Work-conservaion In ecion IV-A, only allocaes minimal bandwidh o each flow, such ha all he flows in a coflow are compleed simulaneously. However, here may be some remaining bandwidh ha can be used o serve more flows. We pursue work-conserving propery by disribuing he remaining bandwidh o flows in o opimize he overall sysem performance. The key poin in disribuing bandwidh is how o deermine he order of flows o preemp he bandwidh. A firs, for he coflows ha have already been scheduled, more bandwidh for any flow in i canno improve is CCT. Therefore, among all coflows, he coflows ha have no been scheduled should have higher prioriy o use he remaining bandwidh; his also helps preven sarvaion. Wihin a coflow, we prefer o allocae more bandwidh o he larger flows han he smaller ones. This is because he flows wih larger raffic volume are more likely o be he boleneck of a coflow, i.e., complee las if all he flows are served by bes-effor delivery. Based on all hese consideraions, we design Algorihm o disribue bandwidh o flows for work conserving purpose. (7) Algorihm : Disribue Remaining Bandwidh : Procedure DisribueBandwidh(Coflows Ω, Bandwidh R) : Non-increasingly sor all he coflows in Ω in erms of heir minimal CCT : for all C Ω do : Non-increasingly sor all he flows in C in erms of flow volume : for all f Ω do : AssignBandwidh(f,R) 7: end for 8: end for 9: end procedure : Procedure AssignBandwidh(Flow f, Bandwidh R) : maxbandwidh : for All he candidae pahs for f, p do : pahbandwidh : for All he links l p do : if R l < pahbandwidh hen : pahbandwidh R l 7: end if 8: end for 9: if pahbandwidh > maxbandwidh hen : maxbandwidh pahbandwidh : end if : end for : reurn maxbandwidh : end procedure In Line, we sor he coflows non-increasingly in erms of heir minimal CCT. In his case, he coflows wih infinie CCT, i.e., ha are no scheduled, can ge higher prioriy o preemp he bandwidh. The procedure AssignBandwidh() assign he bandwidh o corresponding flows. Noe ha he procedure AssignBandwidh() will roue a flow o he pah ha can provide i wih he maximum bandwidh. V. EVALUATION We evaluae hrough a small-scale esbed emulaion as well as large-scale simulaions. chemes o compare: wih. We compare he following schemes : all he flows are roued by ECMP and all of hem fairly compee for bandwidh. cheduling-only (Varys): roues all he flows by ECMP bu schedules hem according o MRTF, which is concepually equivalen o he sae-of-he-ar Varys [9]. Rouing-only: roues all he flows o pursue load balancing bu all he flows should fairly compee for bandwidh. Through comparison wih he las wo schemes, we can inspec he benefis brough by he wo ingrediens of : rouing and scheduling, respecively.

6 .HUQHO 8VHU $OLFDWLRQ (QIRUFHPHQW 7&, HWILOWHU+RRN /LQX[7&+7% )ORZWDEOH 'DHPRQ (QIRUFHPHQW.HUQHORGXOH DFNHWPRGLILHU LRFWO,&GULYHU Fig.. ofware sack of s bandwidh enforcemen. Merics: In his secion, we define he performance of scheme compared o scheme as CCT CCT CCT, where CCT and CCT are he average CCT derived by scheme and scheme, respecively. Wihou declaraion, he performance is compared o baseline scheme. ummary of he main resuls is as follows: Through he experimen on he small-scale leaf-spine esbed (Fig. ), we can see ha 8.% and 8.% of he average CCT can be reduced by, compared o he baseline and rouing-only schemes respecively. Resuls from simulaions repeaedly indicae ha RAPI- ER can reduce he average CCT by up o 79.%,.%, 9.79%, compared o sae-of-he-ar schedulingonly(e.g., Varys [9]), rouing-only, and baseline schemes in differen scenarios. When nework load changes, he performance of is relaively sable; as a comparison, he performances of rouing-only and scheduling-only schemes vary a lo. When iner-coflow arrival inerval is large, consisenly shows very high performance gain. However, even if all he coflows arrive a he same ime (he smalles arrival inerval), can sill achieve.8% performance improvemen in Faree and 9.8% in VL. A. Implemenaion and Tesbed Emulaions Implemenaion: The prooype sysem consiss of he cenral conroller and end hos enforcemen modules. For rouing enforcemen we use he ofware-defined Neworking (DN) echnology o enable explici rouing. For bandwidh enforcemen, we leverage Linux Traffic Conrol (TC) o perform per-flow rae limiing. The archiecure of s bandwidh enforcemen is shown in Fig.. The enforcemen daemon a he user space communicaes wih he kernel module via iocl o manage he flow able. The kernel module, locaing beween TCP/IP sack and TC, inerceps all ougoing packes and modifies nfmark field of socke buffer based on he rules in flow able. The modified packes are hen delivered o TC for rae limiing. We leverage wo-level Hierarchical Token Bucke (HTB) in TC: he roo node classifies packes o heir corresponding Fig.. Tesbed opology. leaf nodes based on nfmark field and he leaf nodes enforce per-flow raes. Tesbed: We build a leaf-spine opology as shown in Fig.. I inerconnecs 9 hoss hrough leaf (ToR) swiches conneced o spine swiches using Gbps links, resuling in a nonblocking fabric. We use Prono 9 8-por Gigabi Eherne swich wih PicO. sysem ha suppors boh Layer / and OpenFlow. Each server has a -core Inel E-.8GHz CPU, 8G memory, GB hard disk and G Eherne NICs. The O of servers is Debian. bi version wih Linux..8. kernel. The CPU, memory or hard disk is no a boleneck in he experimens. We use iperf o generae TCP flows. The base round-rip ime in our esbed is around us. Experimen: In our experimen, we inec coflows ino he nework o evaluae he performance of. As a comparison, we also evaluae he cases of baseline and rouing-only schemes. We do no include scheduling-only because we canno ge he exac flow pahs wih ECMP on he esbed. All he informaion of his experimen is summarized in Table II. I should be noed ha he performance of baseline scheme is averaged by ries, due o he randomness of ECMP. From his experimen, we can see ha can save = 8.% of he average CCT compared o he baseline scheme, and i can reduce he average CCT by = 8.% compared o he rouing-only scheme. Overhead: To make sure ha he overhead of he enforcemen module is negligible, we measured he exra CPU usage inroduced by s enforcemen module. We generaed more han 9Mbps of raffic wih more han flows on a rack server (wih -core Inel E-.8GHz CPU). The exra CPU overhead inroduced was around % (one core) compared wih he case ha s enforcemen module was no used (no rae limiing). The hroughpu remained same in boh cases. Acually, we noe ha, apar from he sofware soluions, some recen hardware soluions [] can also be used o achieve precise rae enforcemen especially a high link speeds, offloading some work from he CPU. B. Large cale imulaions imulaion mehodology: Exising packe-level simulaors such as ns- are no suiable o our case due o heir high overhead []. imilar o [, 9], we develop our own flow-level simulaor. The simulaor accouns for he flow arrival evens and deparure evens, raher han packe sending and receiving

7 Coflow Id# Flow Id# ource Desinaion Volume (GB) M M.7 M M.9 M M9.9 M8 M. M M.9 M7 M M9 M. TABLE II Coflow Compleion Time (s) Rouing-only UMMARY OF TETBED EXPERIMENT: THE AVERAGE CCT OF I 7., ROUTING-ONLY I.7, AND BAELINE I 8.. Average coflow compleion ime (s)..... Performance compared o baseline(%) 8 7 Rouing only (a) Faree 8 8 Coflow widh Rouing only (c) Faree 8 8 Coflow widh Fig.. Average coflow compleion ime (s) 7 Performance compared o baseline(%) Rouing only (b) VL 8 8 Coflow widh (d) VL Rouing only 8 8 Coflow widh The impac of coflow widh. evens, o reduce he simulaion complexiy. I updaes he rae and remaining volume of each flow when an even occurs. To solve linear programming in, we embed he API provided by CPLEX. ino our simulaor. In he simulaions, we use many-o-many communicaion paern wihin a coflow and assume he iner-coflow arrival rae follows a Poisson disribuion. We mainly evaluae aspecs ha may affec he performance of : he widh of a coflow (i.e., he number of flows wihin a coflow), he number of coflows in he nework, and he iner-coflow arrival inerval. For reasonable simulaion ime, we choose -server Faree [] and VL [] as opologies. We also compared he resuls on -server Faree wih ha on 89-server Faree (on which he simulaor runs much slower, over 8 hours for us one ry) and observed similar performance. In he simulaions, each of our resuls is an average of ries. The overall simulaion resuls are shown in Fig. -. In general, we can see ha ouperforms all oher schemes in all scenarios. Impac of coflow widh: In each round of simulaions, we send coflows wih he same widh ino he nework. Fig. shows he simulaion resuls. From his figure, we make he following observaions. Firsly, as shown in Fig. (a) and (b), he absolue average CCT is increased wih he coflow widh. Compared o he baseline scheme, can reduce average CCT by up o 79.% in Faree, and.% in VL. Wihou rouing, he scheduling-only scheme would loss up o 9.%.% =.7% (see Fig. (c) a widh of ) of he performance in Faree. econdly, in Fig. (c) and (d), we observe a rend ha he relaive performance of scheduling-only scheme almos increases wih he coflow widh on boh opologies. The reason is ha when he coflow widh is relaively small, all he coflows are disribued a differen pars of he nework. In his case, coflows are unlikely o compee for bandwidh wih each oher. As a resul, he scheduling does no have much benefi, and rouing-only scheme achieves almos he same performance as. Wih he increase of coflow widh, differen coflows will inerleave wih each oher. Then, scheduling can effecively reduce average CCT by conrolling he flow ransmission raes. Again, in Fig. (c) and (d), he performance of rouingonly scheme increases wih he coflow widh a firs, bu hen decreasing wih i. This is because he flow collision probabiliy (muliple flows are concurrenly acive a he same link) increases wih he flow number in he nework. When he coflow widh is relaively small, rouing scheme can ge good performance as i solves such collision. However, when he coflow widh is relaively large, he rouing-only scheme canno avoid such collision and hence show poor performance. Thirdly, comparing Fig. (c) wih (d), he rouing-only scheme has beer performance in Faree han ha in VL. The reason is ha in Faree, more link-disoined pahs can be found for differen flows if hey are from differen sourcedesinaion pairs, which is no he case in VL. Accordingly, rouing has more opimizaion space o improve he average CCT in Faree. Fourhly, here is an anomaly in VL when he coflow widh is 8 (see Fig. (d)). We should have expeced ha he relaive performance of scheduling-only scheme would increase wih he flow number in he nework. However, his expecaion goes agains he acual simulaion resuls. We noe in Fig. (b) ha large average CCT is caused by many flows in he nework, which makes a large denominaor in he performance definiion. This accouns for he drop. Impac of coflow number: To evaluae how he performance of is influenced by he coflow number in he nework, we fix he coflow widh o be 8. From he resuls in Fig., we make he following observaions. Firsly, as shown in Fig. (a) and (b), he average CCT

8 Average coflow compleion ime (s) 7 Performance compared o baseline(%) Rouing only (a) Faree Coflow number in he nework Rouing only (c) Faree Coflow number in he nework Fig.. Average coflow compleion ime (s) Performance compared o baseline(%) Rouing only (b) VL Coflow number in he nework (d) VL Rouing only Coflow number in he nework The impac of coflow number. Average coflow compleion ime (s) Performance compared o baseline(%) (a) Faree Rouing only....8 Average iner coflow arrival inerval (s)....8 Average iner coflow arrival inerval (s) Fig.. (c) Faree Rouing only Average coflow compleion ime (s) Performance compared o baseline(%) (b) VL Rouing only....8 Average iner coflow arrival inerval (s) Rouing only (d) VL....8 Average iner coflow arrival inerval (s) The impac of iner-coflow number arrival inerval. is increased wih he number of coflows in he nework. always ouperforms rouing-only and scheduling-only schemes obviously. We can see from Fig. (a) ha in Faree he performance of is up o = 8.8% compared o scheduling-only scheme (wih coflows) and = 9.89% compared o rouing-only scheme (wih coflows). econdly, from Fig. (c) and (d), we can see ha, in boh opologies, keeps relaively sable performance wih differen coflow number. The sable performance of comes from is combinaion of rouing and scheduling. When rouing makes less conribuion o wih he increase of coflow number, scheduling can conribue more o compensae he performance loss. Thirdly, we find ha he scheduling-only scheme always ouperforms rouing-only scheme in VL (Fig. (d)), and scheduling-only scheme is more effecive in VL han in Faree (Fig. (c) and (d)). Acually, when here are more coflows compeing wih each oher on he same link, scheduling makes more conribuion o han rouing does. Compared o Faree, here are fewer up links from ToRs in VL, so i is likely ha more flows will inerleave wih each oher in VL han in Faree. Hereby, scheduling is more efficien for VL han for Faree. Impac of iner-coflow arrival inerval: To invesigae he impac of he iner-coflow arrival inerval, we send ou coflows wih he widh of 8 ino he nework, and observe he relaionship beween he sysem performance and arrival inerval of sequenial coflows. Noe ha he larger average iner-coflow arrival inerval indicaes he lower coflow arrival rae. Zero arrival inerval means ha all he coflows arrive a he same ime. We se he larges average iner-coflow arrival inerval o be.8s, since each coflow may complee in a mos s if i monopolizes he nework in our simulaions. From Fig., we make he following observaions. Firsly, as shown in Fig. (a) and (b), he average CCT is decreased wih he increase of average iner-coflow arrival inerval. This is obvious because, as explained above, larger iner-coflow arrival inerval means lower coflow arrival rae. Furhermore, from Fig.(c) and Fig.(d), we find ha wih differen iner-coflow arrival inervals, can reduce CCT by up o 9.79% in Faree and 8.% in VL compared o baseline scheme. Even compared o he rouing-only scheme and scheduling-only scheme, in Faree (Fig.(a)) he performance of can be up o... =.% (when arrival inerval is.s) and = 79.% (when arrival inerval is.s), respecively. econdly, he performance of scheduling-only scheme may firsly increase as he average coflow arrival inerval increases, and hen decrease if he average iner-coflow arrival inerval coninue increasing afer a cerain poin (see Fig.(c) and (d)). When he iner-coflow arrival inerval is small, many coflows should wai for he compleion of oher coflows. Hence, he baseline is large and i resuls in bad performance (see Fig. (a) and (b)). When he iner-coflow arrival inerval is large, he laer coflows may come when he previous coflows almos complee. In his case, he scheduling scheme does no ake effec o reduce he average CCT, since only a few coflows are in nework a he same ime. Thirdly, in Fig.(c) and (d), we can see ha he performance of has he same rend as scheduling-only scheme when he iner-coask arrival inerval is small, while has he same rend as rouing-only scheme when he inercoask arrival inerval is large. This is also because ha scheduling is no effecive o reduce he average CCT when only a few coflows are acive in nework concurrenly, while rouing does no ake effec when oo many coflows in he nework. Takeaways: For reducing average CCT, rouing conribues more when he nework is relaively ligh loaded, since rouing can reduce unnecessary flow collisions. As a comparison, scheduling is more criical when nework load increases, and coflows inerleave wih each oher. The success of is ha i inegraes boh rouing and scheduling, hence always

9 ouperforms oher schemes regardless of he nework saus. VI. RELATED WOR conains wo pars: rouing and scheduling. There is a large specrum of relaed work along eiher rouing or scheduling. We only review some closely relaed ones here. Flow rouing in DCNs: Tradiional raffic engineering soluions inside a daa cener [] or across daa ceners [7, 9] focus on improving he nework resource uilizaion while no reducing he average CCT. They leverage he shor-erm raffic predicabiliy in DCNs o improve he sysem performance. In anoher work, zupdae [] applies o he scenario where some nework componens face failure, while Hedera [] and Due [] focus on how o disribue flows o balance he raffic load in he nework. Relaive o hem, invesigaes how o disribue he flows belonging o he same coflow evenly ino he nework so ha he average CCT can be furher reduced by scheduling. Individual flow scheduling in DCNs: There are also many exising work on opimizing nework uilizaion and reducing average flow compleion ime (FCT) by using scheduling mehods, such as PDQ [] and pfabric []. Boh PDQ and pfabric are flow scheduling schemes o minimize FCT by agging prioriy on he packes. Unforunaely, neiher of hem can be implemened using exising commodiy swiches, and hence hey are no easy o widely deploy. Furhermore, hey do no ake ino accoun he flow dependency semanics and hus are coflow-agnosic. Coflow scheduling in DCNs: Orchesra [8] is perhaps he firs work ha ake he semanics among flow ino accoun when opimizing he flow ransfers in daa cener clusers. Afer ha, he work [7] summarizes he raffic paerns and flow dependency in DCNs and explicily proposes he concep of coflow. Then, recen soluions (e.g., Varys [9] and Barra []) sar o apply he coflow concep (or ask-aware) in heir nework opimizaions, however hey only focus on scheduling while neglecing an indispensable par rouing, which make hese soluions insufficien. VII. CONCLUION is a sysem which opimizes average coflow compleion ime in DCNs by inegraing rouing and scheduling. To he bes of our knowledge, is he firs work ha proposes and proves he posiion ha rouing and scheduling mus be oinly considered for opimizing he average CCT. Through real implemenaion and exensive simulaions, we demonsrae ha works wih exising commodiy swiches and preserves remarkable performance advanages over he scheduling-only or rouing-only soluions. Acknowledgemens This work was suppored in par by HRGC-EC, Naional Basic Research Program of China (97) under Gran CB, CB9, CB, CB78, Huawei Noah s Ark Lab, NFC Fund (7, 7, 79, 77, 9, and 987), and he Fundamenal Research Funds for he Cenral Universiies. REFERENCE [] M. Al-Fares, A. Loukissas, and A. Vahda, A calable, Commodiy Daa Cener Nework Archiecure, IGCOMM Compu. Commun. Rev., vol. 8, no., pp. 7, Aug. 8. [] M. Al-Fares,. Radhakrishnan, B. Raghavan, N. Huang, and A. Vahda, Hedera: Dynamic Flow cheduling for Daa Cener Neworks, in NDI,. [] M. Alizadeh,. Yang, M. harif,. ai, N. Mceown, B. Prabhakar, and. henker, pfabric: Minimal Near-opimal Daacener Transpor, in IGCOMM. [] T. Benson, A. Anand, A. Akella, and M. Zhang, MicroTE: Fine Grained Traffic Engineering for Daa Ceners, in CoNEXT. []. Chen, C. Guo, H. Wu, J. Yuan, Z. Feng, Y. Chen,. Lu, and W. Wu, Generic and Auomaic Address Configuraion for Daa Ceners, in IGCOMM,. []. Chen, A. inglay, A. inghz,. Ramachandranz, L. Xuz, Y. Zhangz, X. Wen, and Y. Chen, OA: An Opical wiching Archiecure for Daa Cener Neworks wih Unprecedened Flexibiliy, in NDI,. [7] M. Chowdhury and I. oica, Coflow: A Neworking Absracion for Cluser Applicaions, in HoNes-XI,. [8] M. Chowdhury, M. Zaharia, J. Ma, M. I. Jordan, and I. oica, Managing daa ransfers in compuer clusers wih orchesra. in IGCOMM. [9] M. Chowdhury, Y. Zhong, and I. oica, Efficien Coflow cheduling wih Varys, in IGCOMM. [] J. Dean and. Ghemawa, MapReduce: implified Daa Processing on Large Clusers, Commun. ACM, vol., no., pp. 7, Jan. 8. [] F. R. Dogar, T. aragiannis, H. Ballani, and A. Rowsron, Decenralized Task-Aware cheduling for Daa Cener Neworks, in IGCOMM. []. Even, A. Iai, and A. hamir, On he complexiy of ime able and muli-commodiy flow problems, in Foundaions of Compuer cience, 97., h Annual ymposium on, Oc 97, pp [] R. Gandhi, H. H. Liu, Y. C. Hu, G. Lu, J. Padhye, L. Yuan, and M. Zhang, Due: Cloud cale Load Balancing wih Hardware and ofware, in IGCOMM. []. Ghemawa, H. Gobioff, and.-t. Leung, The Google File ysem, in OP. [] A. Greenberg, J. R. Hamilon, N. Jain,. andula, C. im, P. Lahiri, D. A. Malz, P. Pael, and. engupa, VL: A calable and Flexible Daa Cener Nework, ACM IGCOMM Compu. Commun. Rev., vol. 9, no., pp., Aug. 9. [] C.-Y. Hong, M. Caesar, and P. B. Godfrey, Finishing Flows Quickly wih Preempive cheduling, in IGCOMM. [7] C.-Y. Hong,. andula, R. Mahaan, M. Zhang, and V. Gill, Achieving High Uilizaion wih ofware-driven WAN, in IGCOMM. [8] M. Isard, M. Budiu, Y. Yu, A. Birrell, and D. Feerly, Dryad: Disribued Daa-parallel Programs from equenial Building Blocks, in Eurosys 7. [9]. Jain, A. umar,. M, J. Ong, L. Pouievski, A. ingh,. Venkaa, J. W, J. Zhou, M. Zhu, J. Zolla, U. Hlzle,. uar, A. Vahda, and G. Inc, B: Experience wih a Globally-Deployed ofware Defined WAN, in IGCOMM. [] H. H. Liu, X. Wu, M. Zhang, L. Yuan, R. Waenhofer, and D. A. Malz, zupdae: updaing daa cener neworks wih zero loss, in ACM IGCOMM. [] G. Malewicz, M. H. Ausern, A. J. Bik, J. C. Dehner, I. Horn, N. Leiser, and G. Czakowski, Pregel: A ysem for Large-scale Graph Processing, in IGMOD. [] A. Munir, G. Baig,. Ireza, I. A. Qazi, F. Dogar, and A. Liu, Friends, no Foes - ynhesizing Exising Daa Cener Transpor raegies, in IGCOMM. [] Y. Peng,. Chen, G. Wang, W. Bai, Z. Ma, and L. Gu, HadoopWach: A Firs ep Towards Comprehensive Traffic Forecasing in Cloud Compuing, in INFOCOM. []. Radhakrishnan, Y. Geng, V. Jeyakumar, A. abbani, G. Porer, and A. Vahda, ENIC: calable NIC for End-hos Rae Limiing, in NDI. [] C. Wilson, H. Ballani, T. aragiannis, and A. Rowron, Beer Never Than Lae: Meeing Deadlines in Daacener Neworks, ACM IGCOM- M Compu. Commun. Rev., vol., no., pp., Aug.. [] M. Zaharia, M. Chowdhury, M. J. Franklin,. henker, and I. oica, park: Cluser Compuing wih Working es, in HoCloud.

Multiprocessor Systems-on-Chips

Multiprocessor Systems-on-Chips Par of: Muliprocessor Sysems-on-Chips Edied by: Ahmed Amine Jerraya and Wayne Wolf Morgan Kaufmann Publishers, 2005 2 Modeling Shared Resources Conex swiching implies overhead. On a processing elemen,

More information

Task is a schedulable entity, i.e., a thread

Task is a schedulable entity, i.e., a thread Real-Time Scheduling Sysem Model Task is a schedulable eniy, i.e., a hread Time consrains of periodic ask T: - s: saring poin - e: processing ime of T - d: deadline of T - p: period of T Periodic ask T

More information

Improvement of a TCP Incast Avoidance Method for Data Center Networks

Improvement of a TCP Incast Avoidance Method for Data Center Networks Improvemen of a Incas Avoidance Mehod for Daa Cener Neworks Kazuoshi Kajia, Shigeyuki Osada, Yukinobu Fukushima and Tokumi Yokohira The Graduae School of Naural Science and Technology, Okayama Universiy

More information

Automatic measurement and detection of GSM interferences

Automatic measurement and detection of GSM interferences Auomaic measuremen and deecion of GSM inerferences Poor speech qualiy and dropped calls in GSM neworks may be caused by inerferences as a resul of high raffic load. The radio nework analyzers from Rohde

More information

The Application of Multi Shifts and Break Windows in Employees Scheduling

The Application of Multi Shifts and Break Windows in Employees Scheduling The Applicaion of Muli Shifs and Brea Windows in Employees Scheduling Evy Herowai Indusrial Engineering Deparmen, Universiy of Surabaya, Indonesia Absrac. One mehod for increasing company s performance

More information

Single-machine Scheduling with Periodic Maintenance and both Preemptive and. Non-preemptive jobs in Remanufacturing System 1

Single-machine Scheduling with Periodic Maintenance and both Preemptive and. Non-preemptive jobs in Remanufacturing System 1 Absrac number: 05-0407 Single-machine Scheduling wih Periodic Mainenance and boh Preempive and Non-preempive jobs in Remanufacuring Sysem Liu Biyu hen Weida (School of Economics and Managemen Souheas Universiy

More information

TSG-RAN Working Group 1 (Radio Layer 1) meeting #3 Nynashamn, Sweden 22 nd 26 th March 1999

TSG-RAN Working Group 1 (Radio Layer 1) meeting #3 Nynashamn, Sweden 22 nd 26 th March 1999 TSG-RAN Working Group 1 (Radio Layer 1) meeing #3 Nynashamn, Sweden 22 nd 26 h March 1999 RAN TSGW1#3(99)196 Agenda Iem: 9.1 Source: Tile: Documen for: Moorola Macro-diversiy for he PRACH Discussion/Decision

More information

PROFIT TEST MODELLING IN LIFE ASSURANCE USING SPREADSHEETS PART ONE

PROFIT TEST MODELLING IN LIFE ASSURANCE USING SPREADSHEETS PART ONE Profi Tes Modelling in Life Assurance Using Spreadshees PROFIT TEST MODELLING IN LIFE ASSURANCE USING SPREADSHEETS PART ONE Erik Alm Peer Millingon 2004 Profi Tes Modelling in Life Assurance Using Spreadshees

More information

Real-time Particle Filters

Real-time Particle Filters Real-ime Paricle Filers Cody Kwok Dieer Fox Marina Meilă Dep. of Compuer Science & Engineering, Dep. of Saisics Universiy of Washingon Seale, WA 9895 ckwok,fox @cs.washingon.edu, [email protected] Absrac

More information

USE OF EDUCATION TECHNOLOGY IN ENGLISH CLASSES

USE OF EDUCATION TECHNOLOGY IN ENGLISH CLASSES USE OF EDUCATION TECHNOLOGY IN ENGLISH CLASSES Mehme Nuri GÖMLEKSİZ Absrac Using educaion echnology in classes helps eachers realize a beer and more effecive learning. In his sudy 150 English eachers were

More information

Chapter 7. Response of First-Order RL and RC Circuits

Chapter 7. Response of First-Order RL and RC Circuits Chaper 7. esponse of Firs-Order L and C Circuis 7.1. The Naural esponse of an L Circui 7.2. The Naural esponse of an C Circui 7.3. The ep esponse of L and C Circuis 7.4. A General oluion for ep and Naural

More information

Trends in TCP/IP Retransmissions and Resets

Trends in TCP/IP Retransmissions and Resets Trends in TCP/IP Reransmissions and Reses Absrac Concordia Chen, Mrunal Mangrulkar, Naomi Ramos, and Mahaswea Sarkar {cychen, mkulkarn, msarkar,naramos}@cs.ucsd.edu As he Inerne grows larger, measuring

More information

Chapter 8: Regression with Lagged Explanatory Variables

Chapter 8: Regression with Lagged Explanatory Variables Chaper 8: Regression wih Lagged Explanaory Variables Time series daa: Y for =1,..,T End goal: Regression model relaing a dependen variable o explanaory variables. Wih ime series new issues arise: 1. One

More information

Morningstar Investor Return

Morningstar Investor Return Morningsar Invesor Reurn Morningsar Mehodology Paper Augus 31, 2010 2010 Morningsar, Inc. All righs reserved. The informaion in his documen is he propery of Morningsar, Inc. Reproducion or ranscripion

More information

Mobile and Ubiquitous Compu3ng. Mul3plexing for wireless. George Roussos. [email protected]

Mobile and Ubiquitous Compu3ng. Mul3plexing for wireless. George Roussos. g.roussos@dcs.bbk.ac.uk Mobile and Ubiquious Compu3ng Mul3plexing for wireless George Roussos [email protected] Overview Sharing he wireless (mul3plexing) in space by frequency in 3me by code PuEng i all ogeher: cellular

More information

Predicting Stock Market Index Trading Signals Using Neural Networks

Predicting Stock Market Index Trading Signals Using Neural Networks Predicing Sock Marke Index Trading Using Neural Neworks C. D. Tilakarane, S. A. Morris, M. A. Mammadov, C. P. Hurs Cenre for Informaics and Applied Opimizaion School of Informaion Technology and Mahemaical

More information

Task-Execution Scheduling Schemes for Network Measurement and Monitoring

Task-Execution Scheduling Schemes for Network Measurement and Monitoring Task-Execuion Scheduling Schemes for Nework Measuremen and Monioring Zhen Qin, Robero Rojas-Cessa, and Nirwan Ansari Deparmen of Elecrical and Compuer Engineering New Jersey Insiue of Technology Universiy

More information

Duration and Convexity ( ) 20 = Bond B has a maturity of 5 years and also has a required rate of return of 10%. Its price is $613.

Duration and Convexity ( ) 20 = Bond B has a maturity of 5 years and also has a required rate of return of 10%. Its price is $613. Graduae School of Business Adminisraion Universiy of Virginia UVA-F-38 Duraion and Convexiy he price of a bond is a funcion of he promised paymens and he marke required rae of reurn. Since he promised

More information

The Transport Equation

The Transport Equation The Transpor Equaion Consider a fluid, flowing wih velociy, V, in a hin sraigh ube whose cross secion will be denoed by A. Suppose he fluid conains a conaminan whose concenraion a posiion a ime will be

More information

Performance Center Overview. Performance Center Overview 1

Performance Center Overview. Performance Center Overview 1 Performance Cener Overview Performance Cener Overview 1 ODJFS Performance Cener ce Cener New Performance Cener Model Performance Cener Projec Meeings Performance Cener Execuive Meeings Performance Cener

More information

TEMPORAL PATTERN IDENTIFICATION OF TIME SERIES DATA USING PATTERN WAVELETS AND GENETIC ALGORITHMS

TEMPORAL PATTERN IDENTIFICATION OF TIME SERIES DATA USING PATTERN WAVELETS AND GENETIC ALGORITHMS TEMPORAL PATTERN IDENTIFICATION OF TIME SERIES DATA USING PATTERN WAVELETS AND GENETIC ALGORITHMS RICHARD J. POVINELLI AND XIN FENG Deparmen of Elecrical and Compuer Engineering Marquee Universiy, P.O.

More information

A Scalable and Lightweight QoS Monitoring Technique Combining Passive and Active Approaches

A Scalable and Lightweight QoS Monitoring Technique Combining Passive and Active Approaches A Scalable and Lighweigh QoS Monioring Technique Combining Passive and Acive Approaches On he Mahemaical Formulaion of CoMPACT Monior Masai Aida, Naoo Miyoshi and Keisue Ishibashi NTT Informaion Sharing

More information

Journal Of Business & Economics Research September 2005 Volume 3, Number 9

Journal Of Business & Economics Research September 2005 Volume 3, Number 9 Opion Pricing And Mone Carlo Simulaions George M. Jabbour, (Email: [email protected]), George Washingon Universiy Yi-Kang Liu, ([email protected]), George Washingon Universiy ABSTRACT The advanage of Mone Carlo

More information

Analysis of Pricing and Efficiency Control Strategy between Internet Retailer and Conventional Retailer

Analysis of Pricing and Efficiency Control Strategy between Internet Retailer and Conventional Retailer Recen Advances in Business Managemen and Markeing Analysis of Pricing and Efficiency Conrol Sraegy beween Inerne Reailer and Convenional Reailer HYUG RAE CHO 1, SUG MOO BAE and JOG HU PARK 3 Deparmen of

More information

MACROECONOMIC FORECASTS AT THE MOF A LOOK INTO THE REAR VIEW MIRROR

MACROECONOMIC FORECASTS AT THE MOF A LOOK INTO THE REAR VIEW MIRROR MACROECONOMIC FORECASTS AT THE MOF A LOOK INTO THE REAR VIEW MIRROR The firs experimenal publicaion, which summarised pas and expeced fuure developmen of basic economic indicaors, was published by he Minisry

More information

GoRA. For more information on genetics and on Rheumatoid Arthritis: Genetics of Rheumatoid Arthritis. Published work referred to in the results:

GoRA. For more information on genetics and on Rheumatoid Arthritis: Genetics of Rheumatoid Arthritis. Published work referred to in the results: For more informaion on geneics and on Rheumaoid Arhriis: Published work referred o in he resuls: The geneics revoluion and he assaul on rheumaoid arhriis. A review by Michael Seldin, Crisopher Amos, Ryk

More information

Strategic Optimization of a Transportation Distribution Network

Strategic Optimization of a Transportation Distribution Network Sraegic Opimizaion of a Transporaion Disribuion Nework K. John Sophabmixay, Sco J. Mason, Manuel D. Rossei Deparmen of Indusrial Engineering Universiy of Arkansas 4207 Bell Engineering Cener Fayeeville,

More information

Acceleration Lab Teacher s Guide

Acceleration Lab Teacher s Guide Acceleraion Lab Teacher s Guide Objecives:. Use graphs of disance vs. ime and velociy vs. ime o find acceleraion of a oy car.. Observe he relaionship beween he angle of an inclined plane and he acceleraion

More information

CHARGE AND DISCHARGE OF A CAPACITOR

CHARGE AND DISCHARGE OF A CAPACITOR REFERENCES RC Circuis: Elecrical Insrumens: Mos Inroducory Physics exs (e.g. A. Halliday and Resnick, Physics ; M. Sernheim and J. Kane, General Physics.) This Laboraory Manual: Commonly Used Insrumens:

More information

The Grantor Retained Annuity Trust (GRAT)

The Grantor Retained Annuity Trust (GRAT) WEALTH ADVISORY Esae Planning Sraegies for closely-held, family businesses The Granor Reained Annuiy Trus (GRAT) An efficien wealh ransfer sraegy, paricularly in a low ineres rae environmen Family business

More information

Chapter 2 Problems. 3600s = 25m / s d = s t = 25m / s 0.5s = 12.5m. Δx = x(4) x(0) =12m 0m =12m

Chapter 2 Problems. 3600s = 25m / s d = s t = 25m / s 0.5s = 12.5m. Δx = x(4) x(0) =12m 0m =12m Chaper 2 Problems 2.1 During a hard sneeze, your eyes migh shu for 0.5s. If you are driving a car a 90km/h during such a sneeze, how far does he car move during ha ime s = 90km 1000m h 1km 1h 3600s = 25m

More information

Why Did the Demand for Cash Decrease Recently in Korea?

Why Did the Demand for Cash Decrease Recently in Korea? Why Did he Demand for Cash Decrease Recenly in Korea? Byoung Hark Yoo Bank of Korea 26. 5 Absrac We explores why cash demand have decreased recenly in Korea. The raio of cash o consumpion fell o 4.7% in

More information

Appendix A: Area. 1 Find the radius of a circle that has circumference 12 inches.

Appendix A: Area. 1 Find the radius of a circle that has circumference 12 inches. Appendi A: Area worked-ou s o Odd-Numbered Eercises Do no read hese worked-ou s before aemping o do he eercises ourself. Oherwise ou ma mimic he echniques shown here wihou undersanding he ideas. Bes wa

More information

SELF-EVALUATION FOR VIDEO TRACKING SYSTEMS

SELF-EVALUATION FOR VIDEO TRACKING SYSTEMS SELF-EVALUATION FOR VIDEO TRACKING SYSTEMS Hao Wu and Qinfen Zheng Cenre for Auomaion Research Dep. of Elecrical and Compuer Engineering Universiy of Maryland, College Park, MD-20742 {wh2003, qinfen}@cfar.umd.edu

More information

DDoS Attacks Detection Model and its Application

DDoS Attacks Detection Model and its Application DDoS Aacks Deecion Model and is Applicaion 1, MUHAI LI, 1 MING LI, XIUYING JIANG 1 School of Informaion Science & Technology Eas China Normal Universiy No. 500, Dong-Chuan Road, Shanghai 0041, PR. China

More information

Appendix D Flexibility Factor/Margin of Choice Desktop Research

Appendix D Flexibility Factor/Margin of Choice Desktop Research Appendix D Flexibiliy Facor/Margin of Choice Deskop Research Cheshire Eas Council Cheshire Eas Employmen Land Review Conens D1 Flexibiliy Facor/Margin of Choice Deskop Research 2 Final Ocober 2012 \\GLOBAL.ARUP.COM\EUROPE\MANCHESTER\JOBS\200000\223489-00\4

More information

What do packet dispersion techniques measure?

What do packet dispersion techniques measure? Wha do packe dispersion echniques measure? Consaninos Dovrolis Parameswaran Ramanahan David Moore Universiy of Wisconsin Universiy of Wisconsin CAIDA [email protected] [email protected] [email protected]

More information

ANALYSIS AND COMPARISONS OF SOME SOLUTION CONCEPTS FOR STOCHASTIC PROGRAMMING PROBLEMS

ANALYSIS AND COMPARISONS OF SOME SOLUTION CONCEPTS FOR STOCHASTIC PROGRAMMING PROBLEMS ANALYSIS AND COMPARISONS OF SOME SOLUTION CONCEPTS FOR STOCHASTIC PROGRAMMING PROBLEMS R. Caballero, E. Cerdá, M. M. Muñoz and L. Rey () Deparmen of Applied Economics (Mahemaics), Universiy of Málaga,

More information

Model-Based Monitoring in Large-Scale Distributed Systems

Model-Based Monitoring in Large-Scale Distributed Systems Model-Based Monioring in Large-Scale Disribued Sysems Diploma Thesis Carsen Reimann Chemniz Universiy of Technology Faculy of Compuer Science Operaing Sysem Group Advisors: Prof. Dr. Winfried Kalfa Dr.

More information

Measuring macroeconomic volatility Applications to export revenue data, 1970-2005

Measuring macroeconomic volatility Applications to export revenue data, 1970-2005 FONDATION POUR LES ETUDES ET RERS LE DEVELOPPEMENT INTERNATIONAL Measuring macroeconomic volailiy Applicaions o expor revenue daa, 1970-005 by Joël Cariolle Policy brief no. 47 March 01 The FERDI is a

More information

Mathematics in Pharmacokinetics What and Why (A second attempt to make it clearer)

Mathematics in Pharmacokinetics What and Why (A second attempt to make it clearer) Mahemaics in Pharmacokineics Wha and Why (A second aemp o make i clearer) We have used equaions for concenraion () as a funcion of ime (). We will coninue o use hese equaions since he plasma concenraions

More information

Hedging with Forwards and Futures

Hedging with Forwards and Futures Hedging wih orwards and uures Hedging in mos cases is sraighforward. You plan o buy 10,000 barrels of oil in six monhs and you wish o eliminae he price risk. If you ake he buy-side of a forward/fuures

More information

Market Liquidity and the Impacts of the Computerized Trading System: Evidence from the Stock Exchange of Thailand

Market Liquidity and the Impacts of the Computerized Trading System: Evidence from the Stock Exchange of Thailand 36 Invesmen Managemen and Financial Innovaions, 4/4 Marke Liquidiy and he Impacs of he Compuerized Trading Sysem: Evidence from he Sock Exchange of Thailand Sorasar Sukcharoensin 1, Pariyada Srisopisawa,

More information

Individual Health Insurance April 30, 2008 Pages 167-170

Individual Health Insurance April 30, 2008 Pages 167-170 Individual Healh Insurance April 30, 2008 Pages 167-170 We have received feedback ha his secion of he e is confusing because some of he defined noaion is inconsisen wih comparable life insurance reserve

More information

A Resource Management Strategy to Support VoIP across Ad hoc IEEE 802.11 Networks

A Resource Management Strategy to Support VoIP across Ad hoc IEEE 802.11 Networks A Resource Managemen Sraegy o Suppor VoIP across Ad hoc IEEE 8.11 Neworks Janusz Romanik Radiocommunicaions Deparmen Miliary Communicaions Insiue Zegrze, Poland [email protected] Pior Gajewski, Jacek

More information

µ r of the ferrite amounts to 1000...4000. It should be noted that the magnetic length of the + δ

µ r of the ferrite amounts to 1000...4000. It should be noted that the magnetic length of the + δ Page 9 Design of Inducors and High Frequency Transformers Inducors sore energy, ransformers ransfer energy. This is he prime difference. The magneic cores are significanly differen for inducors and high

More information

Forecasting, Ordering and Stock- Holding for Erratic Demand

Forecasting, Ordering and Stock- Holding for Erratic Demand ISF 2002 23 rd o 26 h June 2002 Forecasing, Ordering and Sock- Holding for Erraic Demand Andrew Eaves Lancaser Universiy / Andalus Soluions Limied Inroducion Erraic and slow-moving demand Demand classificaion

More information

MTH6121 Introduction to Mathematical Finance Lesson 5

MTH6121 Introduction to Mathematical Finance Lesson 5 26 MTH6121 Inroducion o Mahemaical Finance Lesson 5 Conens 2.3 Brownian moion wih drif........................... 27 2.4 Geomeric Brownian moion........................... 28 2.5 Convergence of random

More information

Chapter 1.6 Financial Management

Chapter 1.6 Financial Management Chaper 1.6 Financial Managemen Par I: Objecive ype quesions and answers 1. Simple pay back period is equal o: a) Raio of Firs cos/ne yearly savings b) Raio of Annual gross cash flow/capial cos n c) = (1

More information

Capacity Planning and Performance Benchmark Reference Guide v. 1.8

Capacity Planning and Performance Benchmark Reference Guide v. 1.8 Environmenal Sysems Research Insiue, Inc., 380 New York S., Redlands, CA 92373-8100 USA TEL 909-793-2853 FAX 909-307-3014 Capaciy Planning and Performance Benchmark Reference Guide v. 1.8 Prepared by:

More information

cooking trajectory boiling water B (t) microwave 0 2 4 6 8 101214161820 time t (mins)

cooking trajectory boiling water B (t) microwave 0 2 4 6 8 101214161820 time t (mins) Alligaor egg wih calculus We have a large alligaor egg jus ou of he fridge (1 ) which we need o hea o 9. Now here are wo accepable mehods for heaing alligaor eggs, one is o immerse hem in boiling waer

More information

Nikkei Stock Average Volatility Index Real-time Version Index Guidebook

Nikkei Stock Average Volatility Index Real-time Version Index Guidebook Nikkei Sock Average Volailiy Index Real-ime Version Index Guidebook Nikkei Inc. Wih he modificaion of he mehodology of he Nikkei Sock Average Volailiy Index as Nikkei Inc. (Nikkei) sars calculaing and

More information

Making a Faster Cryptanalytic Time-Memory Trade-Off

Making a Faster Cryptanalytic Time-Memory Trade-Off Making a Faser Crypanalyic Time-Memory Trade-Off Philippe Oechslin Laboraoire de Securié e de Crypographie (LASEC) Ecole Polyechnique Fédérale de Lausanne Faculé I&C, 1015 Lausanne, Swizerland [email protected]

More information

Impact of scripless trading on business practices of Sub-brokers.

Impact of scripless trading on business practices of Sub-brokers. Impac of scripless rading on business pracices of Sub-brokers. For furher deails, please conac: Mr. T. Koshy Vice Presiden Naional Securiies Deposiory Ld. Tradeworld, 5 h Floor, Kamala Mills Compound,

More information

The naive method discussed in Lecture 1 uses the most recent observations to forecast future values. That is, Y ˆ t + 1

The naive method discussed in Lecture 1 uses the most recent observations to forecast future values. That is, Y ˆ t + 1 Business Condiions & Forecasing Exponenial Smoohing LECTURE 2 MOVING AVERAGES AND EXPONENTIAL SMOOTHING OVERVIEW This lecure inroduces ime-series smoohing forecasing mehods. Various models are discussed,

More information

I. Basic Concepts (Ch. 1-4)

I. Basic Concepts (Ch. 1-4) (Ch. 1-4) A. Real vs. Financial Asses (Ch 1.2) Real asses (buildings, machinery, ec.) appear on he asse side of he balance shee. Financial asses (bonds, socks) appear on boh sides of he balance shee. Creaing

More information

Heuristics for dimensioning large-scale MPLS networks

Heuristics for dimensioning large-scale MPLS networks Heurisics for dimensioning large-scale MPLS newors Carlos Borges 1, Amaro de Sousa 1, Rui Valadas 1 Depar. of Elecronics and Telecommunicaions Universiy of Aveiro, Insiue of Telecommunicaions pole of Aveiro

More information

The Greek financial crisis: growing imbalances and sovereign spreads. Heather D. Gibson, Stephan G. Hall and George S. Tavlas

The Greek financial crisis: growing imbalances and sovereign spreads. Heather D. Gibson, Stephan G. Hall and George S. Tavlas The Greek financial crisis: growing imbalances and sovereign spreads Heaher D. Gibson, Sephan G. Hall and George S. Tavlas The enry The enry of Greece ino he Eurozone in 2001 produced a dividend in he

More information

Usefulness of the Forward Curve in Forecasting Oil Prices

Usefulness of the Forward Curve in Forecasting Oil Prices Usefulness of he Forward Curve in Forecasing Oil Prices Akira Yanagisawa Leader Energy Demand, Supply and Forecas Analysis Group The Energy Daa and Modelling Cener Summary When people analyse oil prices,

More information

Distributed and Secure Computation of Convex Programs over a Network of Connected Processors

Distributed and Secure Computation of Convex Programs over a Network of Connected Processors DCDIS CONFERENCE GUELPH, ONTARIO, CANADA, JULY 2005 1 Disribued and Secure Compuaion of Convex Programs over a Newor of Conneced Processors Michael J. Neely Universiy of Souhern California hp://www-rcf.usc.edu/

More information

Niche Market or Mass Market?

Niche Market or Mass Market? Niche Marke or Mass Marke? Maxim Ivanov y McMaser Universiy July 2009 Absrac The de niion of a niche or a mass marke is based on he ranking of wo variables: he monopoly price and he produc mean value.

More information

Extending IEEE 802.1 AVB with Time-triggered Scheduling: A Simulation Study of the Coexistence of Synchronous and Asynchronous Traffic

Extending IEEE 802.1 AVB with Time-triggered Scheduling: A Simulation Study of the Coexistence of Synchronous and Asynchronous Traffic Exending IEEE 8. wih Time-riggered Scheduling: A Simulaion Sudy of he Coexisence of Synchronous and Asynchronous Traffic Philipp Meyer, Till Seinbach, Franz Korf, and Thomas C. Schmid [email protected],

More information

BALANCE OF PAYMENTS. First quarter 2008. Balance of payments

BALANCE OF PAYMENTS. First quarter 2008. Balance of payments BALANCE OF PAYMENTS DATE: 2008-05-30 PUBLISHER: Balance of Paymens and Financial Markes (BFM) Lena Finn + 46 8 506 944 09, [email protected] Camilla Bergeling +46 8 506 942 06, [email protected]

More information

Analogue and Digital Signal Processing. First Term Third Year CS Engineering By Dr Mukhtiar Ali Unar

Analogue and Digital Signal Processing. First Term Third Year CS Engineering By Dr Mukhtiar Ali Unar Analogue and Digial Signal Processing Firs Term Third Year CS Engineering By Dr Mukhiar Ali Unar Recommended Books Haykin S. and Van Veen B.; Signals and Sysems, John Wiley& Sons Inc. ISBN: 0-7-380-7 Ifeachor

More information

Distributing Human Resources among Software Development Projects 1

Distributing Human Resources among Software Development Projects 1 Disribuing Human Resources among Sofware Developmen Proecs Macario Polo, María Dolores Maeos, Mario Piaini and rancisco Ruiz Summary This paper presens a mehod for esimaing he disribuion of human resources

More information

The option pricing framework

The option pricing framework Chaper 2 The opion pricing framework The opion markes based on swap raes or he LIBOR have become he larges fixed income markes, and caps (floors) and swapions are he mos imporan derivaives wihin hese markes.

More information

CRISES AND THE FLEXIBLE PRICE MONETARY MODEL. Sarantis Kalyvitis

CRISES AND THE FLEXIBLE PRICE MONETARY MODEL. Sarantis Kalyvitis CRISES AND THE FLEXIBLE PRICE MONETARY MODEL Saranis Kalyviis Currency Crises In fixed exchange rae regimes, counries rarely abandon he regime volunarily. In mos cases, raders (or speculaors) exchange

More information

DYNAMIC MODELS FOR VALUATION OF WRONGFUL DEATH PAYMENTS

DYNAMIC MODELS FOR VALUATION OF WRONGFUL DEATH PAYMENTS DYNAMIC MODELS FOR VALUATION OF WRONGFUL DEATH PAYMENTS Hong Mao, Shanghai Second Polyechnic Universiy Krzyszof M. Osaszewski, Illinois Sae Universiy Youyu Zhang, Fudan Universiy ABSTRACT Liigaion, exper

More information

A Note on Using the Svensson procedure to estimate the risk free rate in corporate valuation

A Note on Using the Svensson procedure to estimate the risk free rate in corporate valuation A Noe on Using he Svensson procedure o esimae he risk free rae in corporae valuaion By Sven Arnold, Alexander Lahmann and Bernhard Schwezler Ocober 2011 1. The risk free ineres rae in corporae valuaion

More information

Research on Inventory Sharing and Pricing Strategy of Multichannel Retailer with Channel Preference in Internet Environment

Research on Inventory Sharing and Pricing Strategy of Multichannel Retailer with Channel Preference in Internet Environment Vol. 7, No. 6 (04), pp. 365-374 hp://dx.doi.org/0.457/ijhi.04.7.6.3 Research on Invenory Sharing and Pricing Sraegy of Mulichannel Reailer wih Channel Preference in Inerne Environmen Hanzong Li College

More information

A Load Balancing Method in Downlink LTE Network based on Load Vector Minimization

A Load Balancing Method in Downlink LTE Network based on Load Vector Minimization A Load Balancing Mehod in Downlink LTE Nework based on Load Vecor Minimizaion Fanqin Zhou, Lei Feng, Peng Yu, and Wenjing Li Sae Key Laboraory of Neworking and Swiching Technology, Beijing Universiy of

More information

Supplementary Appendix for Depression Babies: Do Macroeconomic Experiences Affect Risk-Taking?

Supplementary Appendix for Depression Babies: Do Macroeconomic Experiences Affect Risk-Taking? Supplemenary Appendix for Depression Babies: Do Macroeconomic Experiences Affec Risk-Taking? Ulrike Malmendier UC Berkeley and NBER Sefan Nagel Sanford Universiy and NBER Sepember 2009 A. Deails on SCF

More information

PATHWISE PROPERTIES AND PERFORMANCE BOUNDS FOR A PERISHABLE INVENTORY SYSTEM

PATHWISE PROPERTIES AND PERFORMANCE BOUNDS FOR A PERISHABLE INVENTORY SYSTEM PATHWISE PROPERTIES AND PERFORMANCE BOUNDS FOR A PERISHABLE INVENTORY SYSTEM WILLIAM L. COOPER Deparmen of Mechanical Engineering, Universiy of Minnesoa, 111 Church Sree S.E., Minneapolis, MN 55455 [email protected]

More information

Analysis and Design of a MAC Protocol for Wireless Sensor etworks with Periodic Monitoring Applications

Analysis and Design of a MAC Protocol for Wireless Sensor etworks with Periodic Monitoring Applications Analysis and Design of a MAC roocol for Wireless Sensor eworks wih eriodic Monioring Applicaions Miguel A. razo, Yi Qian, Keie u, and Domingo Rodríguez Deparmen of lecrical and Compuer ngineering Universiy

More information

Answer, Key Homework 2 David McIntyre 45123 Mar 25, 2004 1

Answer, Key Homework 2 David McIntyre 45123 Mar 25, 2004 1 Answer, Key Homework 2 Daid McInyre 4123 Mar 2, 2004 1 This prin-ou should hae 1 quesions. Muliple-choice quesions may coninue on he ne column or page find all choices before making your selecion. The

More information

THE FIRM'S INVESTMENT DECISION UNDER CERTAINTY: CAPITAL BUDGETING AND RANKING OF NEW INVESTMENT PROJECTS

THE FIRM'S INVESTMENT DECISION UNDER CERTAINTY: CAPITAL BUDGETING AND RANKING OF NEW INVESTMENT PROJECTS VII. THE FIRM'S INVESTMENT DECISION UNDER CERTAINTY: CAPITAL BUDGETING AND RANKING OF NEW INVESTMENT PROJECTS The mos imporan decisions for a firm's managemen are is invesmen decisions. While i is surely

More information

Modelling and Forecasting Volatility of Gold Price with Other Precious Metals Prices by Univariate GARCH Models

Modelling and Forecasting Volatility of Gold Price with Other Precious Metals Prices by Univariate GARCH Models Deparmen of Saisics Maser's Thesis Modelling and Forecasing Volailiy of Gold Price wih Oher Precious Meals Prices by Univariae GARCH Models Yuchen Du 1 Supervisor: Lars Forsberg 1 [email protected]

More information

Inductance and Transient Circuits

Inductance and Transient Circuits Chaper H Inducance and Transien Circuis Blinn College - Physics 2426 - Terry Honan As a consequence of Faraday's law a changing curren hrough one coil induces an EMF in anoher coil; his is known as muual

More information

Efficient One-time Signature Schemes for Stream Authentication *

Efficient One-time Signature Schemes for Stream Authentication * JOURNAL OF INFORMATION SCIENCE AND ENGINEERING, 611-64 (006) Efficien One-ime Signaure Schemes for Sream Auhenicaion * YONGSU PARK AND YOOKUN CHO + College of Informaion and Communicaions Hanyang Universiy

More information

Statistical Analysis with Little s Law. Supplementary Material: More on the Call Center Data. by Song-Hee Kim and Ward Whitt

Statistical Analysis with Little s Law. Supplementary Material: More on the Call Center Data. by Song-Hee Kim and Ward Whitt Saisical Analysis wih Lile s Law Supplemenary Maerial: More on he Call Cener Daa by Song-Hee Kim and Ward Whi Deparmen of Indusrial Engineering and Operaions Research Columbia Universiy, New York, NY 17-99

More information

UNDERSTANDING THE DEATH BENEFIT SWITCH OPTION IN UNIVERSAL LIFE POLICIES. Nadine Gatzert

UNDERSTANDING THE DEATH BENEFIT SWITCH OPTION IN UNIVERSAL LIFE POLICIES. Nadine Gatzert UNDERSTANDING THE DEATH BENEFIT SWITCH OPTION IN UNIVERSAL LIFE POLICIES Nadine Gazer Conac (has changed since iniial submission): Chair for Insurance Managemen Universiy of Erlangen-Nuremberg Lange Gasse

More information

Chapter 4: Exponential and Logarithmic Functions

Chapter 4: Exponential and Logarithmic Functions Chaper 4: Eponenial and Logarihmic Funcions Secion 4.1 Eponenial Funcions... 15 Secion 4. Graphs of Eponenial Funcions... 3 Secion 4.3 Logarihmic Funcions... 4 Secion 4.4 Logarihmic Properies... 53 Secion

More information

Towards Optimal Capacity Segmentation with Hybrid Cloud Pricing

Towards Optimal Capacity Segmentation with Hybrid Cloud Pricing Towards Opimal Capaciy Segmenaion wih Hybrid Cloud Pricing Wei Wang, Baochun Li, and Ben Liang Deparmen of Elecrical and Compuer Engineering Universiy of Torono Absrac Cloud resources are usually priced

More information

Ecotopia: An Ecological Framework for Change Management in Distributed Systems

Ecotopia: An Ecological Framework for Change Management in Distributed Systems Ecoopia: An Ecological Framework for Change Managemen in Disribued Sysems Tudor Dumiraş 1, Daniela Roşu 2, Asi Dan 2, and Priya Narasimhan 1 1 ECE Deparmen, Carnegie Mellon Universiy, Pisburgh, PA 15213,

More information