Palimpsest: Soft-Capacity Storage for Planetary-Scale Services

Distributed writable storage systems typically provide NFS-like semantics and unbounded persistence for files. We claim that for planetary-scale distributed services such facilities are unnecessary and impose an unwanted overhead in complexity and ease of management. Furthermore, wide-area services have requirements not met by existing solutions, in particular capacity management and a realistic model for billing and charging. We argue for ephemeral storage systems which meet these requirements, and present Palimpsest, an early example being deployed on PlanetLab. Palimpsest is small and simple, yet provides soft capacity, congestion-based pricing, automatic reclamation of space, and a security model suitable for a shared storage facility for wide-area services. It does not “fill up” in the way that an ordinary filing system does. In fact, Palimpsest does not arbitrate resources—storage blocks—between clients at all. 1 The Need for Ephemeral Storage In this paper we address the distributed storage needs of wide-area services designed to run over the kind of shared, planetary-scale computational infrastructure envisioned by the PlanetLab[12], Xenoservers[14], WebOS[18] and Denali[20] projects. So-called “persistent data” in such services consists of the (relatively static) code and data which comprises the service, configuration information, output in the form of log files and transaction records, and system state such as checkpoints and scoreboards. This is a different application area than that addressed by filing systems intended for