Graph connectivity in log-diameter steps using label propagation.
Paul Burkhardt · arXiv (Cornell University) · 2018
The fastest deterministic algorithms for connected components take logarithmic time and perform superlinear work on a PRAM. These algorithms require pointer-chasing operations and are limited to shared-memory systems. Another popular method is `leader contraction' where non-leader vertices are contracted to adjacent leaders. The challenge is to select a constant fraction of leaders that are adjacent to a constant fraction of non-leaders with high probability. Instead we investigate whether simple label propagation can be as efficient as the fastest known algorithms for graph connectivity. Label propagation exchanges representative labels within a component. This is attractive for other models because it is deterministic and does not rely on pointer-chasing, but it is inherently difficult to complete in a sublinear number of steps. We are able to solve the problems with label propagation for graph connectivity. We introduce a simple framework for deterministic graph connectivity in log-diameter steps using label propagation that is easily translated to other computational models. We present new algorithms in PRAM, Stream, and MapReduce. Given a simple, undirected graph $G=(V,E)$ with $n=|V|$ vertices, $m=|E|$ edges, and $D$ diameter, all our algorithms complete in $O(\log D)$ steps without pointer operations. We give the first label propagation algorithms that are competitive with the fastest PRAM algorithms, achieving $O(\log D)$ time and $O((m+n)\log D)$ work with $O(m+n)$ processors. Our main contribution is in Stream and MapReduce models. We give an efficient Stream-Sort algorithm that takes $O(\log D)$ passes and $O(\log n)$ memory, and a MapReduce algorithm taking $O(\log D)$ rounds and $O((m+n)\log D)$ communication overall. These are the first $O(\log D)$-step graph connectivity algorithms in Stream and MapReduce models that are also deterministic and simple to implement.