Hyder - A Transactional Record Manager for Shared Flash

Hyder supports reads and writes on indexed records within classical multi-step transactions. It is designed to run on a cluster of servers that have shared access to a large pool of networkaddressable raw flash chips. The flash chips store the indexed records as a multiversion log-structured database. Log-structuring leverages the high random I/O rate of flash and automatically wear-levels it. Hyder uses a data-sharing architecture that scales out without partitioning the database or application. Each transaction executes on a snapshot, logs its updates in one record, and broadcasts the log record to all servers. Each server rolls forward the log against its locally-cached partial-copy of the last committed state, using optimistic concurrency control to determine whether each transaction commits. This paper explains the architecture of the overall system and its three main components: the log, the index, and the roll-forward algorithm. Simulations and prototype measurements are presented that show Hyder can scale out to support high transaction rates.