Session
Schedule FOSDEM 2020
HPC, Big Data, and Data Science

GraphBLAS: A linear algebraic approach for high-performance graph algorithms

UB5.132
Gabor Szarnyas
There is increasing interest to apply graph analytical techniques to a wide array of problems, many operating on large-scale graphs with billions of edges. While graph algorithms and their complexity is textbook material, efficient implementation of such algorithms is still a major challenge due to a number of reasons. First, the irregular and unstructured nature of graphs leads to a massive amount of random data access, which makes it difficult to use typical caching and parallelization techniques. Second, to optimize their code, developers need to be aware of the nuances of the underlying hardware which, at the very least consists of multiple CPU cores but often also incorporates heterogeneous components such as GPUs or even FPGAs. During the last decade, a number of graph programming models (such as Google's Pregel) have been proposed but most of these focused defining high-level abstractions for distributed execution environments and introduced a significant runtime overhead. A potential approach for defining efficient graph processing algorithms is to exploit the well-known duality of graphs and sparse adjacency matrices, using matrix operations to capture algorithms. Surprisingly, only a few recent research prototypes have used this model with little consensus on the set of necessary building blocks. The GraphBLAS initiative (launched in 2013) aims to define a standard to capture graph algorithms in the language of linear algebra - following the footsteps of the BLAS standard which, starting four decades ago, revolutionized scientific computing by defining constructs on dense matrices. In this talk, I give an overview of the GraphBLAS standard and its key components. First, I illustrate how matrix operations on various semirings correspond to the steps in graph algorithms. I then use these operations to present fundamental graph algorithms such as breadth-first search, shortest paths, and the clustering coefficient. Finally, I demonstrate the scalability of the GraphBLAS-based algorithms with the LDBC Graphalytics benchmark. The presented implementations are available open-source as part of LAGraph, a library built on top of GraphBLAS to demonstrate how to design efficient algorithms in linear algebra.

Intended audience: Developers interested in implementing high-performance graph algorithms.

Expected prior knowledge: Familiarity with linear algebra helps plus but we only use basic concepts such as matrix-matrix multiplication

Additional information

Type devroom

More sessions

2/2/20
HPC, Big Data, and Data Science
Colin Sauze
UB5.132
This talk will discuss the development of a RaspberryPi cluster for teaching an introduction to HPC. The motivation for this was to overcome four key problems faced by new HPC users: The availability of a real HPC system and the effect running training courses can have on the real system, conversely the availability of spare resources on the real system can cause problems for the training course. A fear of using a large and expensive HPC system for the first time and worries that doing something ...
2/2/20
HPC, Big Data, and Data Science
Adrian Woodhead
UB5.132
This presentation will give an overview of the various tools, software, patterns and approaches that Expedia Group uses to operate a number of large scale data lakes in the cloud and on premise. The data journey undertaken by the Expedia Group is probably similar to many others who have been operating in this space over the past two decades - scaling out from relational databases to on premise Hadoop clusters to a much wider ecosystem in the cloud. This talk will give an overview of that journey ...
2/2/20
HPC, Big Data, and Data Science
Félix-Antoine Fortin
UB5.132
Compute Canada provides HPC infrastructures and support to every academic research institution in Canada. In recent years, Compute Canada has started distributing research software to its HPC clusters using with CERN software distribution service, CVMFS. This opened the possibility for accessing the software from almost any location and therefore allow the replication of the Compute Canada experience outside of its physical infrastructure. From these new possibilities emerged an open-source ...
2/2/20
HPC, Big Data, and Data Science
Moritz Meister
UB5.132
Maggy is an open-source framework built on Apache Spark, for asynchronous parallel execution of trials for machine learning experiments. In this talk, we will present our work to tackle search as a general purpose method efficiently with Maggy, focusing on hyperparameter optimization. We show that an asynchronous system enables state-of-the-art optimization algorithms and allows extensive early stopping in order to increase the number of trials that can be performed in a given period of time on ...
2/2/20
HPC, Big Data, and Data Science
Suneel Marthi
UB5.132
The advent of Deep Learning models has led to a massive growth of real-world machine learning. Deep Learning allows Machine Learning Practitioners to get the state-of-the-art score on benchmarks without any hand-engineered features. These Deep Learning models rely on massive hand-labeled training datasets which is a bottleneck in developing and modifying machine learning models. Most large scale Machine Learning systems today like Google’s DryBell use some form of Weak Supervision to construct ...
2/2/20
HPC, Big Data, and Data Science
Frank McQuillan
UB5.132
In this session we will present an efficient way to train many deep learning model configurations at the same time with Greenplum, a free and open source massively parallel database based on PostgreSQL. The implementation involves distributing data to the workers that have GPUs available and hopping model state between those workers, without sacrificing reproducibility or accuracy. Then we apply optimization algorithms to generate and prune the set of model configurations to try.
2/2/20
HPC, Big Data, and Data Science
UB5.132
Predictive maintenance and condition monitoring for remote heavy machinery are compelling endeavors to reduce maintenance cost and increase availability. Beneficial factors for such endeavors include the degree of interconnectedness, availability of low cost sensors, and advances in predictive analytics. This work presents a condition monitoring platform built entirely from open-source software. A real world industry example for an escalator use case from Deutsche Bahn underlines the advantages ...