I attach the patch about this issue.
This patch includes a benchmark class. In this benchmark, for each vertex, the compute function sends a meaningless message into all edges of the vertex. Actually, I intend this benchmark to send messages into random workers. PseudoRandomVertexInputFormat already generates random edges. I employed it.
This benchmark allows users to set the size of message bytes and the number of sending messages per edge. This is because I think they are basic factors to evaluate the behavior and performance of some message delivery system. Besides, users can adjust the number of edges per vertex rather than adjusting the number of sending messages per. It allows users to make the sending pattern either more spread or more skewed.
Anyone can review this?