Reducing communication in parallel graph matching скачать в хорошем качестве

Reducing communication in parallel graph matching

This video shows how communication can be reduced by sending a motivation along with a rejection of a proposal. Furtheemore, futile proposals can be avoided by broadcasting matches. This causes additional communication, in the worst case twice the edge cut, but usually a lot less. On the other hand, the broadcast prevents many useless proposals, and the resulting savings in communication is worthwhile. The video also explains the correspondence between a graph and its adjacency matrix A, and we use this to our advantage by partitioning the matrix A+I, where I is the identity matrix, which expresses that you are not only connected with your friends, but also to yourself. The resulting partitioning for the sparse matrix-vector multiplication with A+I is exactly what we need to minimise the additional communication volume caused by the match broadcasts. We illustrate the partitioning with an example of a social network of 62 dolphins. Finally, we discuss a possible extension to an edge partitioning instead of a vertex partitioning. This may further reduce communication and improve load balance, at the expense of a more complicated algorithm. This video corresponds to Section 5.8 of the book Parallel Scientific Computation: A Structured Approach Using BSP, Second Edition, by Rob H. Bisseling, Oxford University Press, 2020. An expanded set of slides, solutions to the homework questions, and software accompanying the book can all be found on my personal book page: https://webspace.science.uu.nl/~bisse...