Batch-Parallel Compressed Sparse Row: A Locality-Optimized Dynamic-Graph Representation

The default data structure for storing sparse graphs is Compressed Sparse Row (CSR), which enables efficient algorithms but is not designed to accommodate changes to the graph. Since many real-world graphs are dynamic (i.e., they change over time), there has been significant work towards developing dynamic-graph data structures that can support fast algorithms as well as updates to the graph. This paper introduces Batch-Parallel Compressed Sparse Row (BP-CSR), a batch-parallel data structure optimized for storing and processing dynamic graphs based on the Packed Memory Array (PMA). At a high level, Batch-Parallel Compressed Sparse Row extends Packed Compressed Sparse Row (PCSR, HPEC '18), a serial dynamic-graph data structure built on a PMA. However, since the original PCSR runs only on one thread, it cannot take advantage of the parallelism available in multithreaded machines. In contrast, Batch-Parallel Compressed Sparse Row is built on the batch-parallel Packed Memory Array data structure (PPoPP '24) and can support fast parallel algorithms and updates. The empirical evaluation demonstrates that Batch-Parallel Compressed Sparse Row supports fast parallel updates with minimal cost to algorithm performance. Specifically, BatchParallel Compressed Sparse Row performs up to $\mathbf{4 2 0}$ million inserts per second. Across a suite of 10 graph algorithms and 10 input graphs, Batch-Parallel Compressed Sparse Row incurs $1.05 \times$ slowdown on average and about $1.5 \times$ slowdown at most compared to Compressed Sparse Row (CSR), a classical static graph representation. Furthermore, the empirical results show that Batch-Parallel Compressed Sparse Row outperforms existing tree-based and PMA-based dynamic-graph data structures on both algorithms and updates.