112 Large Scale and Big Data
KVs). Then, a series of MapReduce jobs update the state KVs based on the
structure KVs.
• Use of State/Structure KVs. In each iteration, a mapper operating on a
structure KV produces intermediate KVs based the state KVs, and a reducer
updates a state KV. For example, in PageRank, the mapper operating on
each node needs a node’s PageRank score (state KV) and its outgoing neighbors
(structure KV), and the reducer updates a node’s PageRank score (state KV);
In K-means, the mapper operating on each point needs the coordinates of all
centroids (state KVs) and its own coordinate (structure KV), and the reducer
updates a centroid’s coordinate (state KV); In MPI, the mapper 2 operating on
each column of M needs the jth ...