Large networks are useful in a wide range of applications. Sometimes problem instances are composed of billions of entities. Decomposing and analyzing these structures helps us gain new insights about our surroundings. Even if the final application concerns a different problem (such as traversal, finding paths, trees, and flows), decomposing large graphs is often an important subproblem for complexity reduction or parallelization. With even larger instances in applications such as scientific simulation, social networks, or road networks, graph decomposition becomes even more important, multifaceted, and challenging. The seminar is an international forum to discuss recent trends as well as to set new goals and new directions in this research area. The goal of this Dagstuhl Seminar is to bring algorithmic researchers from different communities together who implement, optimize, experiment with algorithms running on large data sets or use techniques from the area frequently, thereby stimulating an exchange of ideas and techniques. The seminar focus will be on graph decomposition. We chose the main topics of our seminar to bring experts together from a wide range of areas – areas that recently have been fruitful in a variety of applications – to tackle some of the most pressing open problems in the area of graph decomposition:

Hardware Design for Dealing with Graphs. Modern processors are optimized for computations that are floating point intensive and have regular memory accesses. Unfortunately, the computations performed by sparse graph algorithms do not benefit from such optimizations. To address this mismatch, new processor and system architectures are being developed. These architectures, exemplified by Intel’s newly announced PUMA architecture, combine highly parallel execution with a word-level accessible global addressable memory. Leveraging the potential of such architectures requires new classes of algorithms.

Beyond Smart Heuristics. In heuristic approaches, improvements in solution quality are often the result of a significant research effort. In recent years, to reduce this effort, develop better heuristics, and ultimately find better solutions, researchers have started developing approaches that use deep neural networks and reinforcement learning in order to learn those heuristics in an end-to-end fashion.

Formulations. Applications that process large sparse data generally have a unique set of optimization requirements for achieving the best performance. Parallelizing such applications on different architectures and/or using different frameworks introduces new performance issues that pertain to these architectures and frameworks. While graphs offer a rich ground for modeling such problems with different requirements, traditional graph decomposition tools may fall short to target those specific issues. This seminar will make more researchers aware of formulations that benefit from the existing state-of-the-art software and encourage them to propose new formulations tailored to their own specific applications.

Scalable Parallel Algorithms for Future Emerging Architectures. Scalable high quality graph partitioning (with quality comparable to sequential partitioners) remains an open problem. With the advent of exascale machines with millions of processors and possibly billions of threads, the situation is further aggravated. Moreover, traditional “flat” partitions of graphs for processing on such machines implies a huge number of blocks. Efficient implementation is also a big issue since complex memory hierarchies and heterogeneity (e.g., GPUs or FPGAs) make the implementation complicated.

Creative Commons BY 4.0

Seher Acer, George Karypis, Christian Schulz, and Darren Strash

• Deepak Ajwani (University College Dublin, IE) [dblp]
• Cevdet Aykanat (Bilkent University - Ankara, TR) [dblp]
• Rob Bisseling (Utrecht University, NL) [dblp]
• Katrin Casel (HU Berlin, DE) [dblp]
• Ümit V. Çatalyürek (Georgia Institute of Technology - Atlanta, US & Amazon Web Services, US) [dblp]
• Cedric Chevalier (CEA - Bruyères-le-Châtel, FR) [dblp]
• Florian Chudigiewitsch (Universität zu Lübeck, DE) [dblp]
• Michael R. Fellows (University of Bergen, NO) [dblp]
• Marcelo Fonseca Faraj (Universität Heidelberg, DE) [dblp]
• Lars Gottesbüren (KIT - Karlsruher Institut für Technologie, DE) [dblp]
• Tobias Heuer (KIT - Karlsruher Institut für Technologie, DE) [dblp]
• George Karypis (University of Minnesota - Minneapolis, US) [dblp]
• Kamer Kaya (Sabanci University - Istanbul, TR) [dblp]
• Jakub Lacki (Google - New York, US) [dblp]
• Johannes Langguth (Simula Research Laboratory - Oslo, NO) [dblp]
• Xiaoye Sherry Li (Lawrence Berkeley National Laboratory, US) [dblp]
• Fredrik Manne (University of Bergen, NO) [dblp]
• Ruben Mayer (Universität Bayreuth, DE) [dblp]
• Johannes Meintrup (THM - Gießen, DE) [dblp]
• Henning Meyerhenke (HU Berlin, DE) [dblp]
• Yosuke Mizutani (University of Utah - Salt Lake City, US) [dblp]
• Francois Pellegrini (University of Bordeaux, FR) [dblp]
• Fabrizio Petrini (Intel Labs - Menlo Park, US) [dblp]
• Frances A. Rosamond (University of Bergen, NO) [dblp]
• Ilya Safro (University of Delaware - Newark, US) [dblp]
• Sebastian Schlag (Apple - Cupertino, US) [dblp]
• Christian Schulz (Universität Heidelberg, DE) [dblp]
• Daniel Seemaier (KIT - Karlsruher Institut für Technologie, DE) [dblp]
• Roohani Sharma (MPI für Informatik - Saarbrücken, DE) [dblp]
• Darren Strash (Hamilton College - Clinton, US) [dblp]
• Blair D. Sullivan (University of Utah - Salt Lake City, US) [dblp]
• Bora Uçar (ENS - Lyon, FR) [dblp]
• Albert-Jan Yzelman (Huawei Technologies - Zürich, CH) [dblp]