Abstract
Abstract: In recent times, the Bulk Synchronous Parallel (BSP) model, which splits the underlying running programs into multiple super-steps, has gain popularity in distributed graph processing systems. However, a high number of message and data exchange in each super-step will create a long epoch of time, often called as network overhead or communication delays. In this paper, we add an additional step in the synchronization phase through decoupling the synchronization into local and global synchronization. This approach moves some of the data and messages exchanged over the network to locally optimized pointers for the next super-step. We prototyped the proposed system and the preliminary evaluation shows decoupled mechanism outperforms baseline system on the average. The system helps eliminate the unnecessary communication between the instances and localize some of the DAG nodes for the next super-step.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have