Parallel and Flow-Based High Quality Hypergraph Partitioning

Balanced hypergraph partitioning is a classic NP-hard optimization problem that is a fundamental tool in such diverse disciplines as VLSI circuit design, route planning, sharding distributed databases, optimizing communication volume in parallel computing, and accelerating the simulation of quantum circuits.
Given a hypergraph and an integer $k$, the task is to divide the vertices into $k$ disjoint blocks with bounded size, while minimizing an objective function on the hyperedges that span multiple blocks.
In this dissertation we consider the most commonly used objective, the connectivity metric, where we aim to minimize the number of different blocks connected by each hyperedge.

The most successful heuristic for balanced partitioning is the multilevel approach, which consists of three phases.
In the coarsening phase, vertex clusters are contracted to obtain a sequence of structurally similar but successively smaller hypergraphs.
Once sufficiently small, an initial partition is computed.
Lastly, the contractions are successively undone in reverse order, and an iterative improvement algorithm is employed to refine the projected partition on each level.
... mehr

An important aspect in designing practical heuristics for optimization problems is the trade-off between solution quality and running time.
The appropriate trade-off depends on the specific application, the size of the data sets, and the computational resources available to solve the problem.
Existing algorithms are either slow, sequential and offer high solution quality, or are simple, fast, easy to parallelize, and offer low quality.
While this trade-off cannot be avoided entirely, our goal is to close the gaps as much as possible.
We achieve this by improving the state of the art in all non-trivial areas of the trade-off landscape with only a few techniques, but employed in two different ways.
Furthermore, most research on parallelization has focused on distributed memory, which neglects the greater flexibility of shared-memory algorithms and the wide availability of commodity multi-core machines.

In this thesis, we therefore design and revisit fundamental techniques for each phase of the multilevel approach, and develop highly efficient shared-memory parallel implementations thereof.
We consider two iterative improvement algorithms, one based on the Fiduccia-Mattheyses (FM) heuristic, and one based on label propagation.
For these, we propose a variety of techniques to improve the accuracy of gains when moving vertices in parallel, as well as low-level algorithmic improvements.
For coarsening, we present a parallel variant of greedy agglomerative clustering with a novel method to resolve cluster join conflicts on-the-fly.
Combined with a preprocessing phase for coarsening based on community detection, a portfolio of from-scratch partitioning algorithms, as well as recursive partitioning with work-stealing, we obtain our first parallel multilevel framework.
It is the fastest partitioner known, and achieves medium-high quality, beating all parallel partitioners, and is close to the highest quality sequential partitioner.

Our second contribution is a parallelization of an n-level approach, where only one vertex is contracted and uncontracted on each level.
This extreme approach aims at high solution quality via very fine-grained, localized refinement, but seems inherently sequential.
We devise an asynchronous n-level coarsening scheme based on a hierarchical decomposition of the contractions, as well as a batch-synchronous uncoarsening, and later fully asynchronous uncoarsening.
In addition, we adapt our refinement algorithms, and also use the preprocessing and portfolio.
This scheme is highly scalable, and achieves the same quality as the highest quality sequential partitioner (which is based on the same components), but is of course slower than our first framework due to fine-grained uncoarsening.

The last ingredient for high quality is an iterative improvement algorithm based on maximum flows.
In the sequential setting, we first improve an existing idea by solving incremental maximum flow problems, which leads to smaller cuts and is faster due to engineering efforts.
Subsequently, we parallelize the maximum flow algorithm and schedule refinements in parallel.

Beyond the strive for highest quality, we present a deterministically parallel partitioning framework.
We develop deterministic versions of the preprocessing, coarsening, and label propagation refinement.
Experimentally, we demonstrate that the penalties for determinism in terms of partition quality and running time are very small.

All of our claims are validated through extensive experiments, comparing our algorithms with state-of-the-art solvers on large and diverse benchmark sets.
To foster further research, we make our contributions available in our open-source framework Mt-KaHyPar.

While it seems inevitable, that with ever increasing problem sizes, we must transition to distributed memory algorithms, the study of shared-memory techniques is not in vain.
With the multilevel approach, even the inherently slow techniques have a role to play in fast systems, as they can be employed to boost quality on coarse levels at little expense.
Similarly, techniques for shared-memory parallelism are important, both as soon as a coarse graph fits into memory, and as local building blocks in the distributed algorithm.