Improving message logging protocols scalability through distributed event logging

Abstract

Message logging is an attractive solution to provide fault tolerance for message passing applications because it is more scalable than coordinated checkpointing. Sender-based message logging is a well known optimization that allows to save messages payload in the sender memory and so only the events corresponding to message receptions have to be logged reliably using an event logger. In existing work on message logging, the event logger has always been considered as a centralized process, limiting message logging protocols scalability. In this paper, we propose a distributed event logger. This new event logger takes advantage of multi-cores processors to be executed in parallel with application processes. It makes use of the nodes' volatile memory to save events reliably. We propose a simple gossip-based dissemination protocol to make application processes aware of new stable events. We evaluated our distributed event logger in the Open MPI library with an optimistic and a pessimistic message logging protocol. Experiments show that distributed event logging improves message logging protocols scalability. © 2010 Springer-Verlag.