1 Data Conversion for Heterogeneous Migration

Migration concerns saving the current computation state, transferring it to remote machines, and resuming execution at the statement following the migration point. Checkpointing concerns saving the computation state to file systems and resuming execution by restoring the computation state from saved files. Although the statetransfer medium differs, migration and checkpointing share the same strategy in state handling. To improve application performance and system resource utilization, they support load balancing, load sharing, data locality optimization, and fault tolerance. The major obstacle preventing migration/checkpointing from achieving widespread use is the complexity of adding transparent migration and checkpointing to systems originally designed to run stand-alone [1]. Heterogeneity further complicates this situation. But migration and checkpointing are indispensable to the Grid [2] and other loosely coupled heterogeneous environments. Thus, effective solutions are on demand. To hide the different levels of heterogeneity, we have developed an application level process/thread migration and checkpointing package, MigThread, which abstracts the computation state up to the language level [3, 4]. For applications written in the C language, states are constructed in the user space instead of being extracted from the original kernels or libraries for better portability across different platforms. A