Abstract:

The MPI3.0 standard contains nonblocking collectives. Nonblocking collectives captures the features of both nonblocking communication and collective communication. Compare with blocking collectives, nonblocking collectives have lower synchronization overhead and can realize the overlap between computing and communication in the program so as to improve the performance. The paper describes different algorithm implementations of broadcast, compares the lowlevel management between nonblocking and blocking broadcast, analyzes the experiment results and finally offers some optimizations.

Key words: MPI;nonblocking broadcast, synchronization overhead;overlap