A Fault-Tolerant Distributed Framework for Asynchronous Iterative Computations

Asynchronous iterative computations (AIC) are common in machine learning and data mining systems. However, the lack of synchronization barriers in asynchronous processing brings challenges for continuous processing while workers might fail. There is no global synchronization point that all workers can roll back to. In this article, we propose a fault-tolerant framework for asynchronous iterative computations (FAIC). Our framework takes a virtual snapshot of the AIC system without halting the computation of any worker. We prove that the virtual snapshot capture by FAIC can recover the AIC system correctly. We evaluate our FAIC framework on two existing AIC systems, Maiter and NOMAD. Our experiment result shows that the checkpoint overhead of FAIC is more than 50 percent shorter than the synchronous checkpoint method. FAIC is around 10 percent faster than other asynchronous snapshot algorithms, such as the Chandy-Lamport algorithm. Our experiments on a large cluster demonstrate that FAIC scales with the number of workers.