| 摘要 |
A method and apparatus may automatically generate a synthetic kernel application for use in optimizing applications on a parallel computing system using protocols such as MPI. The method and system may collect trace communication data from nodes within the parallel computing system running a given application. This trace communication data, which may include function calls and arguments, data size and type, and origin and destination nodes, may be used to calculate processor load along a time axis for each node. A synthetic kernel generator may generate instructions simulating processor load using this communication data. The synthetic kernel generator may also generate instructions to recreate the pattern of node communication with meaningless data, based on the trace communication data. Both types of instructions may then be automatically compiled to form a single synthetic kernel which simulates the behavior of the original application. Other embodiments are described and claimed.
|
