Algorithm for Reducing Overhead of Message Purging in Large-Scale Sender-Based Logging
When attempting to apply sender-based message logging with checkpointing into large-scale and geographically distributed systems, two important things should be reconsidered: reducing the number of messages passing on core networks during its fully message logging and recovery procedures and purging effectively logged messages from their senders volatile memories. This paper presents a novel message purging algorithm to solve the second problem based on our previous work having alleviated the first one. Its first step results in no extra message and forced checkpoint by piggybacking a vector on each sent message. If additional empty buffer space for logging is needed even after the first step has executed, its second step is performed to remove the useful log information each sender maintains while satisfying the consistency condition by using a vector recording its size for every other process.
Keywords: distributed systems, roll-back recovery, message logging, checkpointing, message purging
Download Full-Text
