Xiaoen Ju, Dan Williams, Hani Jamjoom and Kang Shin
USENIX Annutal Technical Conference
Denver, Co, June 2016
Abstract. Multi-version graph processing, where each version
corresponds to a snapshot of an evolving graph, is a
common scenario in large-scale graph
processing. Straightforward application of
existing graph processing systems often yields
suboptimal performance due to high version-switching cost. We present Version Traveler
(VT), a graph processing system featuring fast and
memory-efficient version switching. VT achieves
fast version switching by (i) representing
differences among versions as deltas and (ii)
constructing the next version by integrating the
in-memory graph representation of the current
version with the delta(s) relating the two
versions. Furthermore, VT maintains high
computation performance and memory
compactness. Evaluation using multi-ver- sion
processing workloads with realistic datasets shows
that VT outperforms PowerGraph—running 23x faster
with a 15% memory overhead—and is superior to four
multi-version processing systems.
Bibtex.
@inproceedings{jamjoom-vt-usenix-2016,
author = {Xiaoen and Ju and Dan and Williams and Hani and Jamjoom and Kang and Shin},
title = {{Version Traveler: Fast and Memory-Efficient Version Switching in Graph Processing Systems}},
booktitle = {USENIX Annutal Technical Conference},
address = {Denver, Co},
month = {June},
year = {2016}
}