A Comparison of Two Distributed Disk Systems

As the storage needs of computer applications and users become more sophisticated and increase beyond what can easily be satisfied by a few disk array controllers, aggregating and managing the many disparate components of the storage system become severe problems. Distributed disk systems, which manage collections of disks shared by or partitioned across multiple nodes, may offer a solution to this problem by automating the management of these storage resources. Such systems can automatically tolerate and recover from component failures, gracefully scale in capacity and performance as components are added, allow the storage distributed across multiple nodes to be managed as a single system, and provide useful management abstractions such as virtual disks and snapshots. This paper describes the different architectural and design alternatives embodied in Snappy Disk and Petal, two distributed disk systems that have similar external features but have radically different internal structures, and studies the effect of these differences on the systems’ performance, availability, and manageability. Server0 Server1 Server2 Server3 Client0 Client1 Client2 Client3 General Purpose Network Storage Area Network Server0 Server1 Server2 Server3 Client0 Client1 Client2 Client3 General Purpose Network Figure 1: Shared-Disk Architecture Figure 2: Partitioned-Disk Architecture