Maintenance-Free Global Data Storage

The computing world is experiencing a transition from desktop PCs to connected information appliances, which — like the earlier transition from mainframes to PCs — will profoundly change the way information is used. The variety of devices capable of connecting to the Internet is astounding. Personal Data Assistants (PDAs), cellular phones, and even cars have Internet connectivity. Given the proliferation of connected devices, it’s natural to address their common need for persistent storage by means of an Internet-based, distributed storage system. Such a system would let devices transparently share data and preserve information even when devices are lost or damaged (for a discussion of applications that will benefit from such a storage infrastructure, see the “Applications for Global Storage Systems” sidebar on p. 44). To handle all of the world’s client devices, such a storage system must eventually consist of thousands or millions of Internet-connected servers. Consequently, the greatest obstacle to building and deploying such a system is manageability. Already, single-site data storage systems require carefully orchestrated server configurations, an art understood only by highly paid experts. In a global-scale, distributed storage system, the difficulties increase by many orders of magnitude. Two distinct types of management are readily identified. Storage-level management is the set of operations required to maintain the integrity and access performance of raw data, without reference to the semantics of the information that data represents. Proper storage-level management yields a stable set of bits (or stable store) that can be utilized by applications in a variety of ways. Information-level management, on the other hand, is the set of operations involved in sorting, classifying, and organizing raw data. Consequently, error correction codes are an aspect of storage-level management, while sophisticated indexing is an aspect of information-level management. This article explores mechanisms for storage-level management in OceanStore,1 a global-scale utility infrastructure, designed to scale to billions of users and exabytes of data. A prototype of the OceanStore system is currently under con-