ClusterRAID - architecture and prototype of a distributed fault-tolerant mass storage system for clusters

During the past few years clusters built from commodity off-the-shelf (COTS) components have emerged as the predominant supercomputer architecture. Typically comprising a collection of standard PCs or workstations and an interconnection network, they have replaced the traditionally used integrated systems due to their better price/performance ratio. As paradigms shift from mere computing intensive to I/O intensive applications, mass storage solutions for cluster installations become a more and more crucial aspect of these systems. The inherent unreliability of the underlying components is one of the reasons why no system has been established as a standard storage solution for clusters yet. This thesis sets out the architecture and prototype implementation of a novel distributed mass storage system for commodity off-the-shelf clusters and addresses the issue of the unreliable constituent components. The key concept of the presented system is the conversion of the local hard disk drive of a cluster node into a reliable device while preserving the block device interface. By the deployment of sophisticated erasure-correcting codes, the system allows the adjustment of the number of tolerable failures and thus the overall reliability. In addition, the applied data layout considers the access behaviour of a broad range of applications and minimizes the number of required network transactions. Extensive measurements and functionality tests of the prototype, both stand-alone and in conjunction with local or distributed file systems, show the validity of the concept.