Calculations can naturally be described as graphs in which vertices represe
nt computation and edges reflect data dependencies. By partitioning the ver
tices of a graph, the calculation can be divided among processors of a para
llel computer. However, the standard methodology for graph partitioning min
imizes the wrong metric and lacks expressibility. We survey several recentl
y proposed alternatives and discuss their relative merits. (C) 2000 Elsevie
r Science B.V. All rights reserved.