Item

Abstract

Parallel list ranking is a hard problem due to its extreme degree
of irregularity. Also because of its linear sequential complexity,
it requires considerable effort to just reach speed-up one (break
even). In this paper, we address the question of how to solve the
list-ranking problem for lists of length up to $2 \cdot 10^8$
in practice: we consider implementations on the Intel Paragon,
whose PUs are laid-out as a grid.

It turns out that pointer jumping, independent-set removal and
sparse ruling sets, all have practical importance for current
systems. For the sparse-ruling-set algorithm the speed-up strongly
increases with the number $k$ of nodes per PU, to finally reach
$27$ with 100 PUs, for $k = 2 \cdot 10^6$.