Submitted for publicationCoordinating Adaptations in Self-managing SystemsAn-Cheng Huang, Shang-Wen Cheng, Peter Steenkiste, David Garlan, and Bradley SchmerlSchool of Computer Science, Carnegie Mellon University{pach, zensoul, prs, garlan, schmerl}@cs.cmu.eduAbstract—Self-managing systems are able to dynamicallyadapt to changes in the environment and user needs withouthuman intervention, thus reduce the administrative overheadof system management. Many self-management modules arebeing developed to provide different self-management capabili-ties, and being able to reuse these existing modules will greatlyreduce the necessary efforts for developers of self-managingsystems. However, this requires infrastructure support forplugging different self-management modules into the samesystem and coordination mechanisms for resolving potentialconflicts among adaptations proposed by these modules. In thispaper, we propose a self-managing system architecture thataddresses these two requirements. Using a case study, we showhow multiple self-managing modules can work together on topof a shared infrastructure that provides low-level system accessfunctionalities. We then focus on the issue of coordinatingconflicting adaptations and propose coordination mechanismsbased on observations from the case study.I. INTRODUCTIONThe cost of managing computing systems is becoming amajor issue as computing systems increasingly operate inheterogeneous environments, consist of diverse and distributedelements, and need to deal with changing user needs. Toaddress this problem, an emergent research direction is todevelop self-managing systems that can automatically adaptthemselves to run-time environment changes according tohigh-level policies specified by human developers.Many research projects are developing self-managing sys-tems that focus on different aspects of self-management. Forexample, some may focus on maintaining the overall systemperformance, some tune the run-time parameters of individualsystem elements, and others react to element failures in thesystem. A natural question is: can we use these existingself-managing systems as modules to develop self-managingsystems that can handle multiple aspects of self-management?We believe this can be achieved by (1) providing infrastructuresupport so that existing self-management modules can beplugged into a system, and (2) providing coordination mech-anisms to resolve any potential conflicts among adaptationsproposed by different self-management modules in a system.In this paper, we propose a self-managing system architec-ture that fulfills the above requirements. First, let us describeour model of self-managing systems.A. Self-management modelA self-managing system has been characterized as a col-lection of Autonomic Elements (AEs) [1]. An AE consists ofone Autonomic Manager (AM) and the managed element(s).In such a system, each AM performs adaptations within itsAE according to high-level policies established by developersand to interactions with other AMs.This model can be expanded by adding system-wide self-management. In contrast to element-level adaptations such aschanging the run-time parameters of the managed element,system-wide self-management performs higher-level adapta-tions such as inserting a new AE into the system, removingan old one from the system, and composing a collection of AEsto deliver the desired functionality. Therefore, in this model,there are two types of AMs. An Element-level AM (EAM) isthe AM in the original model and is usually integrated withthe managed element. A System-wide AM (SAM) performssystem-wide adaptations to produce an optimal configurationof AEs according to high-level policies specified by develop-ers.We envision that various SAMs will be developed withdifferent capabilities (e.g., managing the whole system vs.a few elements) and foci (e.g., performance vs. securityconcerns). The developer of a self-managing system can selectthe appropriate SAMs and plug them into the system toadd system-wide self-management capabilities. Examples ofSAMs under development include Libra [2], which finds theglobally optimal configuration for a self-managing system,and Rainbow [3], which adapts to system and environmentchanges.To summarize, a self-managing system may consist ofmultiple EAMs and SAMs. Each AM in the system monitorsthe system state, evaluates a high-level objective specifiedby developers, and acts on the managed system/element(s)accordingly. Such a system can in turn be used as an elementin a higher-level system.B. Application exampleLet us use a video conferencing system to illustrate self-management at both the element and system scopes. As shownin Figure 1, the system is self-managed by two system-wideAMs. SAM1 focuses on dynamically finding a globally opti-mal (minimizing the total latency) configuration of three typesof elements—video conferencing gateway (VGW), handheldproxy (HHP), and end-system multicast [4] proxy (ESMP)—for the users (who have different devices and conferencingapplications). SAM2 has a more targeted focus—when therunning VGW fails, SAM2 finds a lowest-cost VGW to replaceit. On the other hand, at the element level, the three types ofelements are also self-managing, i.e., each element includes anEAM. The ESMPs dynamically adjust the overlay multicastSubmitted for publicationvic/SDRvic/SDRNetMeetingNetMeetinghandheldVGWHHPMulticastoverlayESMPEAMEAMEAMEAMEAMSAM1 SAM2(All elements)ManageDataconnectionFig. 1. A self-managing video conferencing systemtree based on latency and bandwidth performance. The VGWadjust the frame rate according to available bandwidth whenforwarding video streams. The HHP encrypts the data if thehandheld user is connecting from an open wireless network.C. Dimensions of self-managementFrom the above example, we see that the adaptations per-formed by different AMs in the system may potentially conflictwith each other. For example, when the VGW fails, SAM2may find a replacement that is different from what SAM1would use in a globally optimal configuration. Similarly, theESMPs may be adjusting the overlay while SAM1 attemptto use a different set of ESMPs. Therefore, to integrate sucha self-managing system, it is necessary to coordinate theadaptations of different AMs in the system.To better understand when and how adaptations need tobe coordinated, we categorize AMs along the following threedimensions.Granularity: We identify three different granularities of self-management.