1 Christophe Noël De aton1, alma3, Sara Daniel Bouchenak2, Hagimont1, Fabienne drian Mos1 Boyer2, SelfManageable eplicated Servers {Christophe.aton, Grenoble, INI France Sara.Bouchenak, Fabienne.Boyer, 2 University Grenoble, Noel.Depalma, of France Grenoble I Daniel.Hagimont, drian.mos}@inria.fr Grenoble, 3 ING France his selfmanageable presents paper describes bstract reliminary recovering its from case use and with studies server a autonomic middleware replicated failures for systems, automatically solution databases. and and for for Section the 1. automatically servers presented. Introduction according adapting to QoS a cluster requirements of replicated describes are QoS he remainder of the paper is organized as follows. Finally, management requirements 2 section presents scenarios 4 an presents in and of overview which autonomic providing our Jade of conclusions the was failure Jade systems. used middleware management. for and Section ensuring future for 3 makes are replication eplication scalability data their [2][4][7]. and management is replication a availability. wellknown he [6], extremely complexity wo approach and successful difficult ebusiness to of provide such as applications it involves systems service his work. multiple usually requires coordinated the repair manual and help tuning of operations, operators server the management 2. JDE Jade section management middleware first and failure introduces management system, for the autonomic main before in design Jade. discussing principles systems QoS of combined system In this management. paper, skills we database, propose a middlewarebased and operating solution with and proposed may its runtime for use selfmanageable apply with changes, approach different replicated e.g. is and reconfiguration its automatic databases. autonomic generality recovery he systems; on strategies two originality from axes. and to illustrate and Jade is a middleware for the management of autonomic and automatic guarantee of a given quality of failures, First, service of tackle the reconfiguration computing systems. Figure 1 describes the general (QoS). Second, the proposed approach is illustrated here it with architecture the Failure following of Manager. mechanisms, Jade components: and oughly is namely its based speaking, main on the a QoS features control each Manager Jade s loop and with to selfmanagement, application other We replicated implemented software servers. databases; components e.g. Jade, web but a we prototype servers, for show providing that of or the it them may ebusiness proposed apply Second, First, of database the sensors occurrence failure, analysis/decision that or are of a QoS particular responsible requirement components events, for the violation. such detection as. that a middlewarebased solution for selfmanageable systems. with Finally, represent e.g. increasing replicated replacing actuators databases the a number actual failed that upon database of reconfiguration represent high resources load. by the a in new a algorithm, individual cluster one, of relying preliminary for We ensuring then on used experiments QoS databases Jade and with availability replicated illustrate an ebusiness requirements. the usefulness a web cluster. application of Jade Our multitier mechanisms reconfiguration, e.g. necessary allocation of to new implement consists Figure cluster. web 1 illustrates application the use distributed of Jade in with a cluster, an ebusiness node which a end, and two of several components: enterprise servers a web in server the middletier, as a front four replicated database servers as a backend. 2.1. Design principles
automates managed Jade provides Figure a 1. JDE Deployment architecture Manager Cluster makes system. and facilitates use o of two that the other purpose, initial mechanisms deployment Manager the Deployment in Jade: of which he Manager Cluster and Manager the Software is epository. responsible for the cluster which application Manager management the is initially managed component, of the resources free, system or and may is (i.e. may deployed. have nodes) then failed. of be the he used cluster Cluster of by on allocated system/release to provides an application, an I nodes to Jade allocate after deploys use. free Once on nodes those nodes to nodes are an managed the he necessary Software system. software esource components epository that are allows used by business the automatic managed retrieval application. the For software example, resources in case involved of an the software, multitier resources a JBoss J2EE may enterprise [8] a web server MySQL application, software, database the and server used e in pache Manager application epository, Once web nodes server have software been [5]. the and retrieved software from resources allocated the Software necessary by the esource Cluster provided allocated by those nodes. resources managed his is are made by automatically Jade, possible namely due deployed an to I the to I for on an component software remotely based he Jade deploying prototype software was resources implemented nodes. the and resources free model open (e.g. source called MySQL implementation Fractal server [3]. software) Moreover, using a software used a Java encapsulated homogeneously underlying in managed Fractal system components themselves which such as the lifecycle exhibit interface managementrelated (e.g. start/stop operations). interfaces, by herefore, of all whether Jade managed those this management helps components software to provide system actually components, a with generic represent a uniform implementation regardless different view of enables services. legacy bstracting software the Moreover, the development managed systems the software such component as advanced as MySQL Fractal model or components deployment ostgres. used 2.2. dynamic component introspection capabilities provides One for important QoS reconfiguration manager autonomic operations. administration behavior that are resource human consider autonomic intervention utilization Jade behavior is selfoptimization. required. meet which the end classical aims user Selfoptimization needs pattern maximizing with in we no is such deployment balancer standard pattern, QoS and time a distributes infrastructure given and resource a frontend incoming is depicted is proxy replicated requests by acts Figure statically among as a 2. load the In at a replicas. Jade Figure aims 2. autonomously balancing increasing/decreasing among replicas resource efficiently number when the of load adapting replicated increases/decreases. resources utilization used his by has the (i.e. the application preventing effect the o this overbooking). purpose, the QoS manager uses sensors of requests. reconfigure measure the CU the usage he the load QoS system. of the the manager response system. hanks time also hese to the of uses sensors generic applicationlevel actuators can design probe of to Jade, themselves number actuators generic, used since by increasing/decreasing the QoS manager are makes adding/removing Besides of resources sensors components of and actuators, application in the application the is implemented QoS structure. manager the responsible use of an analysis/decision which as selfoptimization notifications (resource increase) for from the algorithm. is implementation sensors required, and, his it increases if component of a the reconfiguration the QoSoriented number receives of is resources available epository deploys adds those nodes. by to contacting retrieve software It then the resources the contacts necessary Cluster the on Manager software the Software new to nodes resources, esource allocate Symmetrically, them to the existing if the application resources structure. allocated to and their a application reconfiguration resources are (i.e. underutilized, to nodes). remove some the QoS replicas manager and performs release requests
performed o If more summarize, llocate by resources the free QoS Figure nodes are manager, required: 3 for describes the which application the are main the following: operations If some necessary Integrate Deploy erform the state required reconciliation software with the other new replicas nodes Unbind if Stop the new replicas to the load balancer. elease resources are underutilized: used. those some the replicas nodes replicas hosting from those the load replicas if no more 2.3. nother Failure autonomic Figure manager administration 3. QoS management in a due repairing replicated Jade to is replication. selfrepair. resource However, In fails, a replicationbased the we service aim behavior remains autonomously system, we available consider replica by the a new managed one. Our system current by goal replacing is to deal with when that failure. managed stop faults. o system this he purpose, proposed as it was the repair prior failure policy to the manager rebuilds occurrence uses the sensors of failed fail probes system; monitor installed these the probes on health the are of nodes implemented the used hosting resources using the heartbeat managed through the techniques. component called he failure the System manager epresentation. also uses he a specific underlying epresentation the and current is used architectural for component failure recovery. structure maintains One of the could a managed representation state that system, System dynamically component model could be used of managed recover crashes, from the system, component introspect failures. and use But encapsulating the that if current a structure node architecture that hosting information replica a is replica of lost; the to that backup is why of the a System component epresentation architecture which is necessary. maintains his a QoS sensors QoS reconfiguration llocate new nodes Cluster Manager Insert replicas Managed resource representation of sense System the the that system epresentation it is (which reflects itself may the replicated is implemented evolve); current to architectural and tolerate as is reliable a snapshot faults. structure in he the actuators, component Besides whole component the the which failure system implements architecture. manager representation, uses the an the autonomic analysis/decision sensors and repair the of sensors epresentation about running behavior. the and, failed It upon receives to node a node retrieve (i.e., notifications failure, software the makes necessary from resources use the of information the that heartbeat that node prior to the failure and their bindings System to esource other allocate resources). epository a new available It to then retrieve contacts node, contacts the necessary Cluster Manager Software software were resources new to operations he this node. System new and performed configuration. redeploys epresentation by the those failure Figure software is then manager. 4 updated resources summarizes according on the Note that Figure the same 4. Failure abstractions management aspect. Furthermore, same system actuators However, for the owing QoS are the aspect used to sensors the to and component differ reconfigure the (components) failure in these abstraction the management two managed and cases. and the reconfiguration to a from component Fractalbased repair reconfiguration model. management capabilities, implementation, capabilities system this repair itself, and of policy therefore i.e. that Jade, can software which benefits used is 3. In implemented and order failure Case to Studies management. validate and tested our several management usecases approach, related we to have QoS Health sensors System representation ecovery reconfiguration llocate new nodes Cluster manager Insert replicas Managed resource
casestudy support servers server). Our (i.e. In first addition and pache experiments we we are web implemented working server involved and on a omcat stateless adding stateful application readwrite replicated ll the for experiments replica recovery. used the benchmark readonly performance application therefore application prototype benchmark environment. similar for to application ebay ubis and is servers. intended an auction [1] It as as management appropriate functionality for present the in validation Jade. of cluster is a 3.1. Stateless his pache experiment QoS replicas management involves dynamic experiments (active one server and webservers idle) implied contained using delivering Jade s identical static dynamic resizing content pages. deployment of and ll a activating cluster servers of active deploy replicated he node pache web pache can load on be an () automatically is idle servers. distributed node. Figure removed by 5 a illustrates proxy an idle that to node the an to can he received be QoS automatically by sensor. in this added, case based is monitoring workload the variations. workload applied Stateful Dynamic Figure replicas clusterresizing, 5. QoS / readonly management illustrated access of stateless in Figure replicas the In a readonly our same experiments, in database this client experiment load. content the s loadbalancer to on an a optimization, all set nodes of DB among (active replicas we preloaded replicated and serving 6, idle). is databases distributed and sensor he database is cjdbc load, [6]. the cjdbc removed in this controller. by case cjdbc based is monitoring the arriving workload DB replicas the from workload variations. the that web can received he be server added QoS by is Stateful We Figure are replicas 6. currently QoS / management readwrite working access of on /O providing stateful replicas technique functionality we in use scenarios leverages with readwrite logging client facilities loads. the of same he previous the jdbc. deployment initial For usecase, each state node operation of all activation, all nodes the have the database the node, the manager DB thus nodes state bringing will preloaded). (as perform it to c In with SQL state reconciliation order the statements synchronization. other to update replicas, operation that the have the new performed log Figure been node file recorded so is used 7 as as to illustrates to part synchronize since replay of the all node last the it activation. depends the number his of the writes time is a during relatively between this state fast time. operation; synchronizations however and it 3.2. We have Figure Failure tested 7. econciliation management Jade s ability experiments to of repair a new running /W replica cluster one a MySQL ubis pache of web database 4 server. omcat application holding he servers omcat scenario the serving application servers in dynamic were which data. connected systems content we had in to a M he Figure casestudy s 8. Failure architecture management is illustrated case study in Figure 8. cjdbc Log cjdbc
when order performance When We under to induced the observe Jade s when system 3 management. consecutive not the was being not evolution under managed omcat Jade of by server management, the Jade application crashes well the as in perceived only essentially manager When remaining automatically the by rendering the system server client the was saturated emulator system recovered managed unavailable. and increased the by the Jade, crashed response dramatically, the servers. failure time either preserve his affected demonstrates system parts availability. Jade s the capacity Note software that to dynamically this architecture assumes repair can crashes a pool is of a available software malfunction nodes exists, and or that the the same cause nodes that and i.e. he be replicas reused presented whose after experiments a restart internal and state involved redeployment. did stateless not need replicas, to be of preserve experiments such, reconciliation observe we the would between repair in operations induce a scenario crashes. functionality consecutive (see We with section plan replicated of DB to Jade 3.1). server perform involving databases. crashes the same state and s 4. Managing Conclusion replicated and systems future is work crashes. enables important particular automatic variations in We large presented reconfiguration enterprise in resource a middleware settings utilization a complex repair that solution and deal of task, server large with that in limiting interventions. consistent clusters By encapsulating of the database, need web for all and costly architectural application and slow entities servers, manual in thus management availability We demonstrated component constraints framework the model, heterogeneous capable QoS Jade management of provides enforcing deployments. a operations QoS uniform and a utilization. functionality and with without database experiments In Jade by addition clusters contrasting that management, involved we for illustrated the preserving evolution automatic a replicated the optimal of failure resizing a system enterprise recovery resource of with web environment implementation deal aspect For with of readwrite future Jade s with of management induced the work scenarios QoS failures. we and in approach failure DB will clusters. managers consolidate will he allow to general better us the optimizations. servers, reconciliation provide a while consistent techniques at the set same of that operations time leverage have valid efficient DBspecific for all state to 5. [1] eferences C.. Specification mza, Gil, J. Marguerite, E. Cecchet, and Implementation K.. ajamani Chanda, of and. Dynamic Cox, W. Zwaenepoel. S. Elnikety, [2] Benchmarks. Characterization http://rubis.objectweb.org BE Logic. IEEE (WWC5), 5th chieving nnual ustin, Workshop Scalability X, US, on Nov. and Workload 2002. High Site [3] vailability http://dev2dev.bea.com/pub/a/2004/01/wls_81_clusterin g.html E. Bruneton,. Coupaye, for EBusiness, and J.B. Stefani. January ecursive 2004. International rogramming http://fractal.objectweb.org/ Dynamic Software (WCO02), Workshop Composition Malaga, on Spain, ComponentOriented with June Sharing. 10, 2002. and 7th [5] [4] ml. B. October http://www.onjava.com/pub/a/onjava/2002/07/10/jboss.ht Cattell, Burke, 2002. S. J. Inscore. Labourey. J2EE Clustering echnology With in JBoss ractice: 3.0. [6] Building echnical Enterprise E. Flexible Cecchet, Business Database Edition. J. Marguerite, pplications earson Clustering Education, W. with Middleware. Zwaenepoel. the 2001. Java 2 FEENIX latform, Conference, Boston, Sessions, M, USENIX EtatsUnis, nnual June 2004. echnical CJDBC: [7] jdbc.objectweb.org/ G. Shachor. omcat Documentation. he pache http://c [8] roject. doc/ Sun (J2EE). Microsystems. http://java.sun.com/j2ee/ http://jakarta.apache.org/tomcat/tomcat3.3 Java 2 latform Enterprise Edition Jakarta