1st High Performance Graph Mining workshop, Sydney, 10 August 2015
http://hdl.handle.net/2117/76366
2020-02-29T08:07:35ZGraSP: distributed streaming graph partitioning
http://hdl.handle.net/2117/76383
GraSP: distributed streaming graph partitioning
Battaglino, Casey; Pienta, Pienta; Vuduc, Richard
This paper presents a distributed, streaming graph parti-
tioner, Graph Streaming Partitioner (GraSP), which makes
partition decisions as each vertex is read from memory, sim-
ulating an online algorithm that must process nodes as they
arrive. GraSP is a lightweight high-performance comput-
ing (HPC) library implemented in MPI, designed to be easily
substituted for existing HPC partitioners such as ParMETIS.
It is the rst MPI implementation for streaming partition-
ing of which we are aware, and is empirically orders-of-
magnitude faster than existing partitioners while providing
comparable partitioning quality. We demonstrate the scala-
bility of GraSP on up to 1024 compute nodes of NERSC's
Edison supercomputer. Given a minute of run-time, GraSP
can partition a graph three orders of magnitude larger than
ParMETIS can.
2015-07-29T11:40:29ZBattaglino, CaseyPienta, PientaVuduc, RichardThis paper presents a distributed, streaming graph parti-
tioner, Graph Streaming Partitioner (GraSP), which makes
partition decisions as each vertex is read from memory, sim-
ulating an online algorithm that must process nodes as they
arrive. GraSP is a lightweight high-performance comput-
ing (HPC) library implemented in MPI, designed to be easily
substituted for existing HPC partitioners such as ParMETIS.
It is the rst MPI implementation for streaming partition-
ing of which we are aware, and is empirically orders-of-
magnitude faster than existing partitioners while providing
comparable partitioning quality. We demonstrate the scala-
bility of GraSP on up to 1024 compute nodes of NERSC's
Edison supercomputer. Given a minute of run-time, GraSP
can partition a graph three orders of magnitude larger than
ParMETIS can.Parallel k nearest neighbor graph construction using tree-based data structures
http://hdl.handle.net/2117/76382
Parallel k nearest neighbor graph construction using tree-based data structures
Rajani, Nazneen; McArdle, Kate; Dhillon, Inderjit S.
Construction of a nearest neighbor graph is often a neces-
sary step in many machine learning applications. However,
constructing such a graph is computationally expensive, es-
pecially when the data is high dimensional. Python's open
source machine learning library Scikit-learn uses k-d trees
and ball trees to implement nearest neighbor graph construc-
tion. However, this implementation is ine cient for large
datasets. In this work, we focus on exploiting these under-
lying tree-based data structures to optimize parallel execu-
tion of the nearest neighbor algorithm. We present parallel
implementations of nearest neighbor graph construction us-
ing such tree structures, with parallelism provided by the
OpenMP and the Galois framework. We empirically show
that our parallel and exact approach is e cient as well as
scalable, compared to the Scikit-learn implementation. We
present the rst implementation of k-d trees and ball trees
using Galois. Our results show that k-d trees are faster when
the number of dimensions is small (2d N); ball trees on
the other hand scale well with the number of dimensions.
Our implementation of ball trees in Galois has almost linear
speedup on a number of datasets irrespective of the size and
dimensionality of the data.
2015-07-29T11:32:22ZRajani, NazneenMcArdle, KateDhillon, Inderjit S.Construction of a nearest neighbor graph is often a neces-
sary step in many machine learning applications. However,
constructing such a graph is computationally expensive, es-
pecially when the data is high dimensional. Python's open
source machine learning library Scikit-learn uses k-d trees
and ball trees to implement nearest neighbor graph construc-
tion. However, this implementation is ine cient for large
datasets. In this work, we focus on exploiting these under-
lying tree-based data structures to optimize parallel execu-
tion of the nearest neighbor algorithm. We present parallel
implementations of nearest neighbor graph construction us-
ing such tree structures, with parallelism provided by the
OpenMP and the Galois framework. We empirically show
that our parallel and exact approach is e cient as well as
scalable, compared to the Scikit-learn implementation. We
present the rst implementation of k-d trees and ball trees
using Galois. Our results show that k-d trees are faster when
the number of dimensions is small (2d N); ball trees on
the other hand scale well with the number of dimensions.
Our implementation of ball trees in Galois has almost linear
speedup on a number of datasets irrespective of the size and
dimensionality of the data.Fast, exact graph diameter computation with vertex programming
http://hdl.handle.net/2117/76371
Fast, exact graph diameter computation with vertex programming
Pennycuff, Corey; Weninger, Tim
In graph theory the diameter is an important topological
metric for understanding size and density of a graph. Unfortunately,
the graph diameter is computationally di cult
to measure for even moderately-sized graphs, insomuch that
approximation algorithms are commonly used instead of exact
measurements. In this paper, we present a new algorithm
to measure the exact diameter of unweighted graphs
using vertex programming, which is easily distributed. We
also show the practical performance of the algorithm in
comparison to other, widely available algorithms and implementations,
as well as the unreliability in accuracy of some
pseudo-diameter estimators.
2015-07-29T08:28:55ZPennycuff, CoreyWeninger, TimIn graph theory the diameter is an important topological
metric for understanding size and density of a graph. Unfortunately,
the graph diameter is computationally di cult
to measure for even moderately-sized graphs, insomuch that
approximation algorithms are commonly used instead of exact
measurements. In this paper, we present a new algorithm
to measure the exact diameter of unweighted graphs
using vertex programming, which is easily distributed. We
also show the practical performance of the algorithm in
comparison to other, widely available algorithms and implementations,
as well as the unreliability in accuracy of some
pseudo-diameter estimators.