Multicoordinated agreement protocols and the log service
Thèse de doctorat : Università della Svizzera italiana, 2008 ; 2008INFO001.
Agreement problems are a common abstraction in distributed systems. They appear when the components of the system must concur on reconfigurations, changes of state, or in lines of action in general. Examples of agreement problems are Consensus, Atomic Commitment, and Atomic Broadcast. In this thesis we investigate these abstractions in the context of the environment in which they will run and the... MoreAdd to personal list
- Agreement problems are a common abstraction in distributed systems. They appear when the components of the system must concur on reconfigurations, changes of state, or in lines of action in general. Examples of agreement problems are Consensus, Atomic Commitment, and Atomic Broadcast. In this thesis we investigate these abstractions in the context of the environment in which they will run and the applications that they will serve; in general, we consider the asynchronous crash-recovery model. The goal is to devise protocols that explore the contextual information to deliver improved availability. The correctness of our protocols holds even when the extra assumptions do not. In the first part of this thesis we explore the following property: messages broadcast in small networks tend to be delivered in order and reliably. We make three contributions in this part. The first contribution is to turn known Consensus algorithms that harness this ordering property to reach agreement in the crash-stop model into practical protocols. That is, protocols that tolerate message losses and recovery after crashes, efficiently. Our protocols ensure progress even in the presence of failures, if spontaneous ordering holds frequently. In the absence of spontaneous ordering, some other assumption is required to cope with failures. The second contribution of this thesis is to generalize one of our crash-recovery consensus protocols as a ``multicoordinated'' mode of a hybrid Consensus protocol, that may use spontaneous ordering or failure detection to progress. Compared to other protocols, ours provide improved availability with no price in resilience. The third contribution is to employ this new mode to solve Generalized Consensus, a problem that generalizes a series of other agreement problems and, hence, is of much practical interest. Moreover, we considered several aspects of solving this problem in practice, which had not been considered before. As a result, our Generalized Consensus protocol features graceful degradation, load balancing, and is parsimonious in accessing stable storage. In the second part of this thesis we have considered agreement problems in wide area networks organized hierarchically. More specifically, we considered a topology that is commonplace in the data centers of large corporations: groups of nodes, with large-bandwidth low-latency links connecting the nodes in the same group, and slow and limited links connecting nodes in different groups. In such environments, latency is clearly a major concern and reconfiguration procedures that render the agreement protocol momentarily unavailable must be avoided as much as possible. Our contribution here is in avoiding reconfigurations and improving the availability of a collision fast agreement protocol. That is, a protocol that can reach agreement in two intergroup communication steps, irrespectively to concurrent proposals. Besides the use of a multicoordinated approach, we employed multicast primitives and consensus to restrict some reconfigurations to within groups, where they are less expensive. In the last part of this thesis we study the problem of terminating distributed transactions. The problem consists of enforcing agreement among the parties on whether to commit or rollback the transaction and ensuring the durability of committed transactions. Our contribution in this topic is an abstract log service that detaches the termination problem from the processes actually performing the transactions. The service works as a black box and abstracts its implementation details from the application utilizing it. Moreover, it allows slow and failed resource managers be re-started on different hosts without relying on the stable storage of the previous host. We provide two implementations of the service, which we evaluated experimentally.