Hybrid Checkpointing for Iterative Processing in BSP-Based Systems

Abstract

Distributed iterative processing exists in various application scenarios including large-scale graph analytics and machine learning. Many systems employ bulk synchronous parallel (BSP) model to synchronize the iterations. In these BSP-based systems, the long iterative processing time in distributed environments makes the fault-tolerance crucial. Most BSP-based systems write a checkpoint in either blocking strategy or unblocking strategy to achieve fault-tolerance. However, the blocking strategy involves a checkpointing overhead in failure-free cases, whereas the unblocking strategy also incurs a recovery cost if the BSP-based system has not completed checkpointing in failure cases. Motivated by the trade-off between blocking and unblocking checkpointing, we aim to choose different checkpointing strategy when checkpoint is required during iterative processing, in order to reduce the whole execution time. In particular, we propose a checkpointing choice problem, i.e., how to choose the strategy to minimize the execution time. The challenge is to make a choice during runtime without future information. To address this problem, we provide a hybrid checkpointing, which heuristically chooses either blocking or unblocking checkpointing based on cost evaluation. Our experiments on Giraph, a typical BSP-based system, show that hybrid checkpointing outperforms blocking and unblocking checkpointing.