Hostng Vrtual Machnes on Dstrbuted Datacenters Chuan Pham Scence and Engneerng, KyungHee Unversty, Korea pchuan@khu.ac.kr Jae Hyeok Son Scence and Engneerng, KyungHee Unversty, Korea sonaehyeok@khu.ac.kr Nguyen H.Tran Scence and Engneerng, KyungHee Unversty, Korea nguyenth@khu.ac.kr Choong Seon Hong Engneerng, KyungHee Unversty, Korea cshong@khu.ac.kr Mnh N.H. Nguyen Scence and Engneerng, KyungHee Unversty, Korea mnhnhn@khu.ac.kr ABSTRACT Almost of cloud servces nowadays are bult at top geographcally dstrbuted nfrastructure for better relablty and performance. These cloud provders need an effcent method to control and drect user workload to sutable datacenter, dependng on many factors such as: network traffc cost, operaton cost budget, energy consumpton, etc. In the vrtual machne placement problem, current works manly focus on the effcency of packng vrtual machnes nto servers and gnore the dstrbuted scenaro of datacenters. In ths paper, we consder the problem of placng vrtual machnes to host applcatons on a shared resource pool based on dstrbuted cloud platforms. We formulate the problem of hostng vrtual machnes on dstrbuted datacenters as an optmzaton problem, and propose a dstrbuted framework DHC that can dynamcally drect workload between datacenters to maxmze total utlty of entre datacenters. We also conduct many case studes to valdate our method, and evaluate ts effectveness and practcalty, usng real-workload traffc traces. The smulaton results show that our algorthm can dynamcally optmze the total utlty, dependng on varous workload of users. Categores and Subect Descrptors C.2.3 [Network Operatons]: Resoure management. Keywords Optmzaton theory, Vrtual machne placement problem, server, queung theory, network traffc, delay.. INTRODUCTION Correspondng author. Permsson to make dgtal or hard copes of all or part of ths work for personal or classroom use s granted wthout fee provded that copes are not made or dstrbuted for proft or commercal advantage and that copes bear ths notce and the full ctaton on the frst page. Copyrghts for components of ths work owned by others than ACM must be honored. Abstractng wth credt s permtted. To copy otherwse, or republsh, to post on servers or to redstrbute to lsts, requres pror specfc permsson and/or a fee. Request permssons from permssons@acm.org. IMCOM 6, January 04-06, 206, Danang, Vet Nam c 206 ACM. ISBN 978--4503-442-4/6/0... $5.00 DOI: http://dx.do.org/0.45/2857546.2857633 Cloud computng s becomng a rsng paradgm n the nformaton and communcatons technology ndustry today. Many cloud servces are deployed on multple datacenters that are located n dfferent regon for mprovng performance and relablty. The term dstrbuted systems (such as dstrbuted computng, dstrbuted databases, etc.) generally refer to that scenaro. Fg. presents the archtecture of dstrbuted cloud computng, where many datacenters cooperate to share the resource pool. Dstrbuted cloud solutons are very useful to cloud provders to enable vrtual network servces that span a wde geographcal area. It supports more securely. Because all data are not n the same place. It wll not be damageable too much n any dsaster and wll easly recover n small amount of tme. Ths technque also does not need a large resource pool such as network capacty, CPU, memory, and storage at datacenters. The cloud servce s spread out everywhere and wll be more lkely to be close from where users are accessng t. Furthermore, wth small capacty, dstrbuted cloud computng requres less electrcty almost cuts off power consumpton from ar condtonng Moreover, n servng workload, when the number of requests from users n cloud servces ncreases sgnfcantly, dstrbuted clouds not only mtgate the sufferng from massve connecton but also mprove load balancng and fault tolerant n cloud servces. Secondly, clent requests come from wde area. Current approaches often dstrbute requests to closest datacenters to reduce the latency and bandwdth of network. However, consderng on entre network, requests should balance between datacenters to reduce the operaton cost n some specfc datacenters, havng hgh workloads. In addton, each datacenter has dfferent resource management polces, such as revenue, dfference amount of avalable resources, power consumpton, etc. For example, geographcal load balancng and energy cost polcy are managed ndependently by each datacenter. It causes poor performance and hgh costs n many cases []. Workng on power consumpton of datacenters, the authors of [5] proposed a method that dynamcally scales the number of actve servers n datacenter dependng on workloads to reduce the power consumpton n datacenters. However, they dd not propose for the dstrbuted cloud datacenters where many provders mplement on large-scale geo-
Data center Gateway Data center Data center Fgure : The VM allocaton on dstrbuted datacenters. graphcally dstrbuted nfrastructure nowadays. Desgnng an effcent cooperatve manner that can share the pool of resources and obtan maxmum revenue s stll a potental problem n dstrbuted datacenters. In ths paper, we formulate the vrtual machne (VM) placement problem n dstrbuted datacenters and propose a framework that can maxmze the socal-welfare of all datacenters. By usng the theory of dual decomposton method, we dvde the master problem nto multple sub-problems, whch can easly solve at each datacenter. The rest of the paper s organzed as follows. Secton II represents the related work. The system model and problem formulaton s represented n Secton III. In Secton IV, we solve the optmal problem by decomposton method. In Secton V, we descrbe the experment envronment and smulaton results. Fnally, we conclude our work n Secton VI. 2. RELATED WORKS Interest n VM placement n datacenters, there are several papers focus on optmzng the total utlty of all servers n a datacenter [2, 3, 4]. Usng dfferent methods, they successfully pack VMs nto servers at one datacenter, but n the dstrbuted scenaro, these methods should be consdered more factors that really affect to allocaton scheme, such as power consumpton at each datacenter, network latency, etc. Moreover, the users often submt a bundle of resource demand that ncludes a lst of VMs. These VMs have nsde relatonshps that cannot arbtrarly host n dfferent servers or datacenters. In another trend, accountng the operaton cost of datacenters, many proposals [5, 6, 7, 8] focus on the centralze method to optmal resource management, reducng carbon footprnt problem, or load balancng n the centralze datacenter. The authors n [9] proposed an autonomc method that provdes resources for servce hostng platforms. That method consdered wth two components: global decson module and local decson module. Ths archtecture can automatcally dstrbute the tasks nto approprate servers and optmze the provsonng resource. However, ths s a method that apples nto the centralzed datacenter. It s also the NP-hard problem n VM placement, whch s dffcult to fnd out the optmal soluton. Mnghong Ln et al [5] proposed a method that dynamcally scales the number of actve servers n datacenters dependng on workloads. Ths method s a good way to reduce the power consumpton n datacenters. But they dd not propose for the dstrbuted cloud. We consder smlar approach but extend nto dstrbuted scenaro of datacenters, where the complexty of user demand s complcated and there are many factors mpact to the effcency of the controllng scheme of datacenters. 3. SYSTEM MODEL AND PROBLEM FOR- MULATION 3. System modellng In ths secton, we show an archtecture of our method that llustrates how to partal executon nto each datacenter and how to control provsonng resource n dstrbuted datacenters. Each datacenter conssts of a set of actve physcal servers. Each server can host multple vrtual machnes through a hypervsor. To reduce the power consumpton, the number of actve servers at each datacenter can be changed dependng on workloads. We defne X = (x, x 2,..., x,..., x N ) s a vector of actve servers, whch x s the number of actve servers at datacenter. We consder a dscrete tme model, where n each slot the cloud servce provder (as a gateway n Fg. ) receves amount of tasks A = (a, a 2,..., a,..., a M ). Each task a s assocated wth specfc performance goals specfed n a servce-level agreement (SLA) contract. An a can embed an arbtrary applcaton topology, whch requres a lst of VMs, whch s called a cluster. To reduce network transfer, a cluster s often hosted at one datacenter. To model the provson resource n cloud, we consder each VM belong to class k. For example, Amazone EC2 defnes a lst of nstants that clents can choose properly to ther applcaton. We defne that a s the VM allocaton vector of task a = (n, n 2,..., n k,..., n c), where n k s the number of VMs of class k attrbuted to task a, and c s maxmum number of class. Each task a requres amount of resource R (such as: memory, CPU unt, storage, etc.) that can be calculated as follows: R s = c k n k I s k, s S. () where I s k s amount of resource type s of nstance k, and S s a set of consdered resources. Resource constrant. To host task a, a datacenter has to turn on a number of servers that have suffcent resource. We consder the datacenter has homogeneous servers that can calculate the total avalable resource based on the number of actve server x as follows C(x ) = βx, =,..., N, (2) where β s the vector resource of one server (ncludng CPU, memory, storage, etc.). Therefore, the total resource hostng at datacenter cannot exceed the capacty C(x ) as follows H R s C s (x ), =,..., N, s S, (3) where H s an ndcatng bnary varable, H = f the task a s hosted at datacenter, otherwse, H = 0. Further, each task a can host totally at one datacenter to guarantee the qualty of servces of task. The constrant s represented as follows H =, =,..., M. (4)
Delay constrant. Whle mprovng the qualty of servces (QoS), we consder that the allocaton of datacenters and demand assgnment must satsfy a set of constrants, such as: demand constrant and Servce Level Agreement (SLA) performance constrant. We defne σ be the demand arrval rate from the gateway assgned to datacenter (σ = N = H), the demand constrant ensures that all demands are satsfed: σ = D, (5) where D s the average demand arrval rate orgnated from the gateway. In addton, we defne d as the network latency between the gateway and datacenter. At datacenter, we assume that the arrvng demand σ s equally splt among x actve servers. From the queueng model M/M/ [], the queueng delay between the gateway to a server can be computed as: L = µ σ x, k =,..., N, (6) where µ s the servce rate of each server. The constrant of delay s defned as: µ σ x + d d max, =,..., N, (7) where d max s the maxmum delay that the servce provder tres to acheve. By defnng the constant α = µ We can rewrte the constran (7) as: d max d, =,..., N. (8) x σ α, =,..., N. (9) Budget constrant. In some specal cases, the total operaton cost at one datacenter s qute hgh. Especally, when the emergency demand response [2] comes n, the datacenters have to cut-off the power consumpton by server turnng on/off or workload mgraton, etc. We defne the budget B for each datacenter as the hghest operaton cost at datacenter. Ths constrant depends on the number of actve servers at datacenter as follows γx B, =,..., N, (0) where γ s the prce that converts the operaton cost to a monetary term. 3.2 The dstrbuted datacenter optmzaton problem We am at fndng the allocaton scheme for each task a whle maxmzng a global utlty value U, whch s expressed as a weghted sum of the applcaton-provded resource-level utlty functons and operatng cost functon [9]. The resourcelevel utlty functon u for task a s defned as u (R ) that presents the user satsfacton based on the amount of allocated resource (user satsfacton s used nterchangeably n ths paper). The dstrbuted cloud model can be formulated as: max U = H u γx () H R s C s (x ), =,..., N, s S, (2) H =, =,..., M, (3) x σ α, =,..., N, (4) γx B, =,..., N, (5) H = [0, ], =,..., M, =,..., N, (6) var. {H, x}. (7) Problem ()-(7) s an mx-nteger lnear program (MILP), whch s solved by a hgh complexty algorthm. Wth thousands of servers n a datacenter, we can further relax the nteger H as contnuous varables such that ths problem s tractable. For ease of notaton, we name the problem ()-(7) as DHC (Dstrbuted Hostng datatacenter). 4. DISTRIBUTED CLOUD COMPUTING EM- BEDDING DECOMPOSITION ARCHITEC- TURE By relaxng the nteger constrant, the optmzaton problem DHC s essentally a lnear programmng (LP), and can be solved n polynomal tme. However, ths requres a central coordnator whch gather nformaton from all datacenters. Further, the complexty of solvng the LP also ncreases sgnfcantly when the problem sze scales up. Thus, we are motvated to develop dstrbuted solutons n whch the mappng nodes teratvely solve DHC problem. Partcularly, the relaxed problem DHC has follows max H u γx + υ ( H ) (8) H R s C s (x ), =,..., N, s S, (9) x σ α, =,..., N, (20) γx B, =,..., N, (2) where υ s Lagrange multpler assocated wth constrant (3). The dea of decomposton problem above s to separate the master problem nto many sub-problem, n whch each datacenter tres to optmal the resource allocaton ndependently. The gateway wll collect nformaton from all datacenters and control resource prce for all the subproblems. Each datacenter, the subproblem s as follows:
00 000 Utlty 650 550 500 450 B =5 B =7 B =20 500 0 5 0 5 20 25 30 Tme slots 350 0 5 0 5 20 25 30 35 Tme slots Fgure 2: Sample of workload for 3 observed tmeslots. Utlty 300 200 00 Fgure 3: DHC. 0 0 20 30 40 50 60 Iteratons Evaluaton of the fast convergence of max H u γx + υ H (22) H R s C s (x ), s S, (23) x σ α, (24) γx B. (25) At each teraton, gven varable H and υ, N subproblems are solved smultaneously to obtan the optmal varable x. We apply Dantzg-Wolfe decomposton technque to fnd the optmal assgnment varable H and υ n each teraton by solvng the master problem as descrbed n [9, chap.2]. The obectve of the master problem monotoncally ncreases n each teraton and converges to optmal soluton of the orgnal LP. In general, a gateway can nstantate a set of polces at the master problem, after recevng requests, dctatng the order n whch the varables need to be optmzed and obtan the mappng matrx H and vector X. The gateway receves all nformaton from datacenters and updates H (based on the formula n [9, chap.2]) to calculate for the next step. Ths process runs untl convergence. 5. SIMULATION AND NUMERICAL In ths secton, our goal s n two-folds: Frst, we mplement our model to evaluate the performance of DHC. Second, we seek to llustrate the effcent of our model to n- Fgure 4: Evaluaton of utlty at datacenter. crease revenue n the context of realstc workloads of Google cluster-usage traces [3] as shown n Fg. 2. 5. Settngs To capture the locaton dversty of the cloud nfrastructure, we set the number of datacenter N = 0. For each datacenter, we consder three types of VM nstances and one resource type (c = 3, s = ). Each datacenter conssts of homogeneous servers whch have equally servce rate µ = 0.2. In term of the monetary weghts, we set γ = 0.0. The delay d s set randomly n range 2 to 5ms and d max s set randomly n range 0 to 20ms. 5.2 Results Convergence. Accordng to settngs above, we execute DHC to evaluate the convergence of our method. The result n Fg. 3 matches wth the characterstcs of dstrbuted algorthm DHC, where the convergence occurs after 40 teratons. Usng dstrbuted method, the complexty of DHC s reduced sgnfcantly n large-scale of practcal envronment. The huge number of servers n dstrbuted datacenters s broken down and separately solved at each datacenter. Total utlty. To evaluate the effcency of DHC, we measure the total utlty at a specfc datacenter (Datacenter ). By settng dfferent budged B, we measure the alternatve of utlty at datacenter as shown n Fg. 4. Dependng on the gven budget, datacenter tres to maxmze ts utlty. However, when the cost to serve the arrval workload exceeds the budget, the datacenter wll forward the workload to other datacenters. For example, at tme slot 0, wth budget B = 7, the datacenter does not receve the arrval workload. Meanwhle, wth hgher budget B = 20, the fluctuaton of utlty reflects the dynamc workload at ths datacenter. In another measurement, we take nto account total utlty of entre network, as shown n Fg. 5. The total utlty of dstrbute datacenters follows workload traffc that dynamcally controls the assgnment workload at each datacenter and balances between operaton cost and budget. However, n Fg. 5b, network traffc cost s so hgh, each datacenter tres to serve workload at ts local datacenter and does not forward to others. In duraton tme slot 4 to 9, snce total operaton costs of all dstrbuted datacenter come to the lmtaton of the budged B, all datacenters are frozen. Ths shows the flexble n servng workload of each datacenter to guarantee ts beneft. However, n practcal envronment, at that tme the controller of datacenters wll run the backup servces, or specal polces to guarantee the workload be
00 000 Total utlty 0 0 0 5000 0 0 5 0 5 20 25 30 35 3000 Tme slots (a) Evaluaton of total utlty of dstrbute datacenters wth normal traffc cost. 00 000 Utlty 0 0 0 5000 0 0 5 0 5 20 25 30 35 3000 Tme slots (b) Evaluaton of total utlty of dstrbute datacenters wth hgh traffc cost. Fgure 5: Comparson between normal network traffc and hgh network traffc. executed as least one datacenter, as mentoned n [4]. 6. CONCLUSION In ths paper, we study hostng vrtual machnes n dstrbuted datacenter. In vew that exstng studes on VM placement problem have been largely solated to date and resulted n uncoordnated management n dstrbuted scenaro. We propose a coordnated approach to enablng sharng workload between datacenter to maxmze total utlty of entre dstrbuted datacenter. We present a dstrbute soluton, called DHC, that can optmally control workload, delay and budget at each datacenter. We also conduct many case studes to valdate our method, showng that DCH can dynamcally control workload at all dstrbuted datacenter. The analyss and smulaton show that our proposed system s adaptable and can be appled to the schedulng functon of data centers. As a future work, we want to apply wdely our model n more practcal envronment, focus deeply about resource constrants of VMs that ncreases the complexty of DHC problem. 7. ACKNOWLEDGMENTS Ths research was supported by the MSIP, Korea, under the G-ITRC support program (IITP-205-R682-5-000) supervsed by the IITP. Dr. CS Hong s the correspondng author. 8. REFERENCES [] Mad Hababa and Saed Gorgn. A revew on modern dstrbuted computng paradgms: Cloud computng, ungle computng and fog computng. CIT. Journal of Computng and Informaton Technology, 22(2):69 84, 204. Utlty Utlty [2] Huandong Wang, Yong L, Yng Zhang, and Depeng Jn. Vrtual machne mgraton plannng n software-defned networks. arxv preprnt arxv:42.4980, 204. [3] Faben Hermener, Xaver Lorca, Jean-Marc Menaud, Glles Muller, and Jula Lawall. Entropy: a consoldaton manager for clusters. In Proceedngs of the 2009 ACM SIGPLAN/SIGOPS nternatonal conference on Vrtual executon envronments, pages 4 50. ACM, 2009. [4] F Ma, F Lu, and Z Lu. Mult-obectve optmzaton for ntal vrtual machne placement n cloud data center. J. Infor. and Computatonal Scence, 9(6), 202. [5] Mnghong Ln, Adam Werman, Lachlan LH Andrew, and Eno Thereska. Dynamc rght-szng for power-proportonal data centers. IEEE/ACM Transactons on Networkng (TON), 2(5):378 39, 203. [6] Gong Chen, Wenbo He, Je Lu, Suman Nath, Leondas Rgas, Ln Xao, and Feng Zhao. Energy-aware server provsonng and load dspatchng for connecton-ntensve nternet servces. [7] Hrshkesh Amur, James Cpar, Varun Gupta, Gregory R Ganger, Mchael A Kozuch, and Karsten Schwan. Robust and flexble power-proportonal storage. In Proceedngs of the st ACM symposum on Cloud computng, pages 27 228. ACM, 200. [8] Peter Bodk, Mchael Paul Armbrust, Kevn Cann, Armando Fox, Mchael Jordan, and Davd A Patterson. A case for adaptve datacenters to conserve energy and mprove relablty. Unversty of Calforna at Berkeley, Tech. Rep. UCB/EECS-2008-27, 2008. [9] Hen Nguyen Van, Frederc Dang Tran, and Jean-Marc Menaud. Autonomc vrtual resource management for servce hostng platforms. In Proceedngs of the 2009 ICSE Workshop on Software Engneerng Challenges of Cloud Computng, pages 8. IEEE Computer Socety, 2009. [0] Pelle Jakovts, Satsh Narayana Srrama, and Ila Kromonov. Stratus: A dstrbuted computng framework for scentfc smulatons on the cloud. In Hgh Performance Computng and Communcaton & 202 IEEE 9th Internatonal Conference on Embedded Software and Systems (HPCC-ICESS), 202 IEEE 4th Internatonal Conference on, pages 053 059. IEEE, 202. [] Sheldon M Ross. Introducton to probablty models. Access Onlne va Elsever, 2006. [2] Mohamed H Albad and EF El-Saadany. Demand response n electrcty markets: An overvew. In IEEE power engneerng socety general meetng, volume 2007, pages 5, 2007. [3] Google cluster-usage traces. URL http://code.google.com/p/googleclusterdata/. [4] Shaole Ren Nguyen Tran, Chuan Pham and Choong Seon Hong. Coordnated energy management for emergency demand response n mxed-use buldngs. 205.