Performance Isolation in Cloud-Based Big Data Architectures

Cloud-based big data systems usually have many different tenants that require access to the server's functionality. In a nonisolated cloud system, the different tenants can freely use the resources of the server. Hereby, disruptive tenants who exceed their limits can easily cause degradation of performance of the provided services for other tenants. To ensure performance demands of the multiple tenants and meet fairness criteria, various performance isolation approaches have been introduced including artificial delay, round robin, blacklist, and thread pool. Each of these performance isolation approaches adopts different strategies to avoid the performance interference in case of multiple concurrent tenant needs. In this paper, we propose a framework and a systematic approach for performance isolation in cloud-based big data systems. To this end, we present an architecture design of cloud-based big data system and discuss the integration of feasible performance isolation approaches. We evaluate our approach using PublicFeed, a social media application that is based on a cloud-based big data platform.