Graph pattern mining (GPM) extracts subgraph structures (e.g., motifs and cliques) from large graphs, but traditional vertex-centric graph computing frameworks struggle with the explosive growth of intermediate embeddings and redundant isomorphism checks. Building on the first GPM system, Arabesque, this thesis presents a systematic survey and an empirical study of modern GPM systems. We then conduct a comparative evaluation across six real-world graphs (CiteSeer, Mico, YouTube, LiveJournal, Orkut, and Twitter20) and three representative mining tasks: Motif Counting (MC), Clique Finding (CF), and Triangle Counting (TC) on large graphs. The experiments measure runtime together with CPU and memory usage. Results show clear specialization across systems: Peregrine is the most efficient for MC on small to large graphs; Sandslash (and, to a lesser extent, Pangolin) is strongest on CF and is the only tested system to finish 4-clique finding on the billion-edge Twitter20 graph, and Arya delivers state-of-the-art TC on large graphs, completing Orkut and Twitter20 in fractions of a second, far faster than CPU-centric baselines. Overall, pattern-aware exploration, decomposition (with or without sampling), and heterogeneous execution are the dominant factors behind scalable performance, while exhaustive embedding enumeration remains the primary bottleneck. These findings provide a unified view of the GPM systems landscape and practical guidance for building the next generation of scalable graph mining systems.