Memory - Mapped Transactions

Memory-mapped transactions combine the advantages of both memory mapping and transactions to provide a programming interface for concurrently accessing data on disk without explicit I/O or locking operations. This interface enables a programmer to design a complex serial program that accesses only main memory, and with little to no modification, convert the program into correct code with multiple processes that can simultaneously access disk. I implemented Libxac, a prototype for an efficient and portable system supporting memorymapped transactions. Libxac is a C library that supports atomic transactions on memory-mapped files. Libxac guarantees that transactions are serializable, and it uses a multiversion concurrency control algorithm to ensure that all transactions, even aborted transactions, always see a consistent view of a memory-mapped file. Libxac was tested on Linux, and it is portable because it is written as a user-space library, and because it does not rely on special operating system support for transactions. With Libxac, I was easily able to convert existing serial, memory-mapped implementations of a B-tree and a cache-oblivious B-tree into parallel versions that support concurrent searches and insertions. To test the performance of memory-mapped transactions, I ran several experiments inserting elements with random keys into the Libxac B-tree and Libxac cache-oblivious B-tree. When a single process performed each insertion as a durable transaction, the Libxac search trees ran between 4% slower and 67% faster than the B-tree for Berkeley DB, a high-quality transaction system. Memory-mapped transactions have the potential to greatly simplify the programming of concurrent data structures for databases. Thesis Supervisor: Charles E. Leiserson Title: Professor of Computer Science and Engineering Thesis Supervisor: Bradley C. Kuszmaul Title: Research Scientist