Mechanisms for improving ZooKeeper Atomic Broadcast performance

Abstract

Coordination services are essential for building higher-level primitives that are often used in today’s data-center infrastructures, as they greatly facilitate the operation of distributed client applications. Examples of typical functionalities offered by coordination services include the provision of group membership, support for leader election, distributed synchronization, as well as reliable low-volume storage and naming. To provide reliable services to the client applications, coordination services in general are replicated for fault tolerance and should deliver high performance to ensure that they do not become bottlenecks for dependent applications. Apache ZooKeeper, for example, is a well-known coordination service and applies a primary-backup approach in which the leader server processes all state-modifying requests and then forwards the corresponding state updates to a set of follower servers using an atomic broadcast protocol called Zab. Having analyzed state-of-the-art coordination services, we identified two main limitations that prevent existing systems such as Apache ZooKeeper from achieving a higher write performance: First, while this approach prevents the data stored by client applications from being lost as a result of server crashes, it also comes at the cost of a performance penalty. In particular, the fact that it relies on a leader-based protocol, means that its performance becomes bottlenecked when the leader server has to handle an increased message traffic as the number of client requests and replicas increases. Second, Zab requires significant communication between instances (as it entails three communication steps). This can potentially lead to performance overhead and uses up more computer resources, resulting in less guarantees for users who must then build more complex applications to handle these issues. To this end, the work makes four contributions. First, we implement ZooKeeper atomic broadcast, extracting from ZooKeeper in order to make it easier for other developers to build their applications on top of Zab without the complexity of integrating the entire ZooKeeper codebase. Second, we propose three variations of Zab, which are all capable of reaching an agreement in fewer communication steps than Zab. The v variations are built with restriction assumptions that server crashes are independent and a server quorum remains operative at all times. The first variation offers excellent performance but can only be used for 3-server systems; the other two are built without this limitation. Then, we redesigned the latest two Zab variations to operate under the least-restricted Zab fault assumptions. Third, we design and implement a ZooKeeper coin-tossing protocol, called ZabCT which addresses the above concerns by having the other, non-leader server replicas toss a coin and broadcast their acknowledgment of a leader’s proposal only if the toss results in an outcome of Head. We model the ZabCT process and derive analytical expressions for estimating the coin-tossing probability of Head for a given arrival rate of service requests such that the dual objectives of performance gains and traffic reduction can be accomplished. If a coin-tossing protocol, ZabCT is judged not to offer performance benefits over Zab, processes should be able to switch autonomously to Zab. We design protocol switching by letting processes switch between ZabCT and Zab without stopping message delivery. Finally, an extensive performance evaluation is provided for Zab and Zab-variant protocols.