Programmed graph reduction based on function granularity and its parallel implementation

This paper examines closely the implementation problems-code execution, global memory access, and graph distribution of the programmed graph reduction machine based on function granularity, and proposes their solutions. Our function-grained graph reduction machine enables tasks to reduce the graph nodes that represent function applications. Each processing element executes multiple tasks concurrently, and manages itself with its own image of the operating system. In the implementation of this machine, we represented a machine code as a macro instruction that calls a predefined program, and devised DSM system that makes PEs access a virtual global memory but delays sending the update to the remote page until execution of a machine code is terminated successfully. We used the dipstick scheduling strategy, which combines both the partition and the distribution strategies to allocate graph nodes to PEs. We also showed the evaluation results that are experimented on the prototype machine