A Data Placement Scheme with Time-Related Model for P2P Storages

Abstract

Maintaining desired data availability while minimizing costs is the primary challenge in designing P2P storages. Data placement schemes and data availability calculation methods are two key design choices. Prior work has used the random placement and steady-state availability analysis model widely. We find this approach has two drawbacks: (1) it ignores the time-related property and greatly underestimates data availability in the short term and (2) it ignores differences in peers' availability and uses a mean value, which also leads to the underestimation of data availability. Both of these drawbacks influence the efficiency of storage and bandwidth resource usage. Also, this approach can not ensure all the objects achieve their expected availability due to the second drawback. This paper develops a different data placement scheme called similar-MTTF-MTTR placement, which takes into consideration differences in pees' dynamic characteristics. Under this scheme, this paper also presents a fine-gained analysis model for short-term data availability calculation. As a result, our scheme improves the efficiency of data maintaining significantly. We validate this by comparing with random placement under three different dynamic environments: a wide-area system (PlanetLab), a corporative environment (desktop personal computers at Microsoft Corporation), and a file-sharing system (Maze). The results show that our scheme can save the storage resource usage significantly under all environments.