Advances in hardware have enabled many long-running applications to execute entirely in main memory. With the emergence of cloud computing, thousands of machines could be made available to deploy such applications with lowered operational and maintenance costs. While achieving substantially better performance, these applications have encountered new challenges in achieving fault tolerance; i.e., to ensure durability in the event of a crash. In addition, many of these applications, such as massively multiplayer online games, main-memory OLTP systems, main-memory search engine and deterministic transaction processing systems, must sustain extremely high update rates - often hundreds of thousands of updates per second. They also demand extremely high throughput (e.g. scientific simulation) or low latency (e.g. massively multiplayer online games). To support these demanding requirements, these applications have increasingly turned to database techniques. In this dissertation, we propose an approach to provide fault tolerance for main-memory applications without introducing excessive overhead or latency spikes. First, we evaluate the applicability of existing checkpoint recovery techniques developed for main-memory DBMS. We use massively multiplayer online games (MMOs) as our motivating example. In particular, we show how to adapt consistent checkpointing techniques developed for main-memory databases to MMOs. Furthermore, we provide a thorough simulation model and evaluation of six recovery strategies. Based on our results, we argue that not all state-of-the-art checkpoint recovery techniques are equally suited for low-latency and high-throughput applications such as MMOs. These algo- rithms either use locks or large synchronous copy operations, which hurt throughput and latency, respectively. Next, we take advantage of frequent points of consistency in many of these applications to develop novel checkpoint recovery algorithms that trade additional space in main memory for significantly lower overhead and latency. Compared to previous work, our new algorithms do not require any locking or bulk copies of the application state. Our experimental evaluation shows that one of our new algorithms attains nearly constant latency and reduces overhead by more than an order of magnitude for low to medium update rates. Additionally, in a heavily loaded main-memory transaction processing system, it still reduces overhead by more than a factor of two. Finally, we present BRRL, a library for making distributed main-memory applications fault tolerant. BRRL is optimized for cloud applications with frequent points of consistency that use data-parallelism to avoid complex concurrency control mechanisms. BRRL differs from existing recovery libraries by providing a simple table abstraction and using schema information to optimize checkpointing.
展开▼
