Atomic transfer for distributed systems

Building applications and information systems increasingly means dealing with concurrency and faults stemming from distribution of system components. Atomic transactions are a well-known method for transferring the responsibility for handling concurrency and faults from developers to the software's execution environment, but incur considerable execution overhead. This dissertation investigates methods that shift some of the burden of concurrency control into the network layer, to reduce response times and increase throughput. It anticipates future programmable network devices, enabling customized high-performance network protocols. We propose Atomic Transfer (AT), a distributed algorithm to prevent race conditions due to messages crossing on a path of network switches. Switches check request messages for conflicts with response messages traveling in the opposite direction. Conflicting requests are dropped, obviating the request's receiving host from detecting and handling the conflict. AT is designed to perform well under high data contention, as concurrency control effort is balanced across a network instead of being handled by the contended endpoint hosts themselves. We use AT as the basis for a new optimistic transactional cache consistency algorithm, supporting execution of atomic applications caching shared data. We then present a scalable refinement, allowing hierarchical consistent caches with predictable performance despite high data update rates. We give detailed I/O Automata models of our algorithms along with correctness proofs. We begin with a simplified model, assuming static network paths and no message loss, and then refine it to support dynamic network paths and safe handling of message loss. We present a trie-based data structure for accelerating conflict-checking on switches, with benchmarks suggesting the feasibility of our approach from a performance stand-point.