On the Design and Scalability of Distributed Shared-Memory Databases

To scale-out databases today, partitioning data across several database instances is the common technique. This approach is typically referred to as shared-nothing architecture. In this paper, we propose a new alter- native architecture design for distributed relational databases based on two fundamental principles: We logically decouple query processing and transaction management from data storage and we share data across all query processing nodes. The combination of these design choices provides operational exibility, a property that enables cost-ecient database in- frastructures and that becomes increasingly important with the growing market penetration of cloud computing. As a drawback, sharing data among several database nodes adds substantial synchronization overhead. We present techniques for ecient data access and concurrency control to counter this eect. These techniques are enabled by recent hardware trends such as in-memory storage and low-latency networking and pro- vide scalability characteristics that top state-of-the-art shared-nothing databases