João Luís (2011)

MSC Student

in

MSc dissertation: TxBtrfs — A Transactional Snapshot-based File System  
Period: October 2010 — September 2011
Grade: 18/20
Project: Synergy-VM project (funded by the National Science Foundation, PI: Prof. João Lourenço)
Papers: InForum'11 (link).

Several decades ago, the file system was the container of choice for large bulks of related information, kept in hundreds of files, and relying on applications specifically created to handle them. These configurations weren't scalable and could easily become difficult to maintain, leading to the development and adoption of Database Management Systems (DBMS). These systems, capable of efficiently handling vast amounts of data, allowed heavy concurrency without requiring the programmer to deal with concurrency-control mechanisms, by encapsulating operations within transactions.

The properties of Transactions rapidly became an object of desire by many, and efforts to bring them to general-purpose programming environments began. In recent years there have been breakthroughs in bringing the transactional semantics to memory, using Software Transactional Memory (STM), providing abstractions to concurrency-control on the application-level. However, STM failed to meet some expectations, specially regarding I/O operations, forcing the abstraction to go deeper in the system: directly to the file system.

In this document we shall discuss file systems in general, their properties and common structure, although focusing in those with transactional or versioning capabilities. Later on, we will present our proposed enhancement of an existing Linux file system (Btrfs), in order to offer transactional semantics to applications, while detecting potential conflicts between concurrent flows of execution and reconciling their changes whenever possible.