Our research objective is providing checkpoint-based fault tolerance for message passing processes on grids. As a research result, we've implemented MPICH-GF which stands for fault-tolerant MPICH-G2. MPICH-GF consists of hierarchical process managers and MPICH-GF library. Hierarchical managers control parallel processes, monitor failures and recover failed processes automatically. MPICH-GF library includes dynamic process management, checkpoint library and checkpointing and message logging protocols. Current MPICH-GF implementation supports coordinated checkpointing protocol, in which processes generate checkpoint files simultaneously after coordination and all of processes rollback on a single failure. We've succeeded in testing Nas Parallel Benchmark applications with MPICH-GF. MPICH-GF guarantees no message-loss in global checkpoint and supports non-blocking message transfer. While most of previous related works adopt indirect message transfer method, in MPICH-GF processes transfer message directly to target processes. Indepedent checkpointing with message logging and high available process manager implementation are in progress.