Self Tuning With Self Confidence∗

Recent research on managing complex computing systems has focused on the autonomic computing vision: Systems should manage themselves according to high level administrative goals [5]. As an example, system components should monitor the environment and self-tune to meet quality of service expectations, without requiring manual intervention in selecting concrete configuration options or in coordinating the reconfiguration process. In a distributed system, the usual approach is to have several alternative protocols for each key distributed system function, each serving a different performance tradeoff. Depending on the assessment of the environment and on the policy set by the administrator, the ideal protocol is chosen and configured. The result should be a simple feedback control loop that provides the desired self-tuning capability. This approach has however several disadvantages. First, having multiple implementations, that may be seldom used, increases the complexity of the system and has implications in the reliability of the software. Second, the switchover mechanism is itself complex and often requires distributed agreement, thus introducing a large overhead and not seldom the need to “stop the world” while in progress. Also, by switching among a limited set of options, the system can only cope with a limited number of scenarios. And finally, practical implementations often rely on a centralized coordinator component, which is in itself an obstacle to dependability. In short, all these issues combined make current systems fall short of the autonomic computing vision. In this paper we show how two fundamental building blocks for dependable distributed systems – consensus and reliable multicast – can be extended to support autonomic principles. In fact, we point out that existing protocols [6, 1] support decentralized decisions and fine grained adaptation with a single algorithm thus surpassing the impairments pointed above. Furthermore, the correctness of the algorithms is always ensured no matter what adaptive decisions are taken. Finally, we speculate on realizing actual sys-