DFS Cycle Detection on Pregel Model

The patent segments the cycle detection process into multiple iterations, where each iteration detects cycles of a specific length. Instead of attempting to detect all cycles simultaneously (which would require storing all paths in memory), the algorithm divides the problem into n-1 iterations, each handling cycles of length k. This segmentation allows the system to process large graphs with limited memory by focusing on one cycle length at a time.

The patent employs preliminary action by pre-assigning unique identifiers (IDs) to vertices and preparing the graph structure before cycle detection begins. Each vertex is pre-configured with its ID, and the system pre-establishes the iteration framework. This preliminary setup enables the iterative DFS process to proceed efficiently without requiring excessive memory during the actual detection phase, as the basic structure is already in place.

The patent inverts the traditional BFS approach by using depth-first search (DFS) instead. In BFS, messages are propagated level-by-level across the graph, requiring many messages to reach distant vertices. In contrast, DFS follows paths deeply before backtracking, allowing the algorithm to detect cycles with fewer message transmissions. This inversion of the search strategy fundamentally reduces the message complexity from O(V+E) in BFS to a more efficient pattern suited for cycle detection.

The patent introduces dynamics by making the search strategy adaptive through iterations. Each iteration k focuses on detecting cycles of length k, dynamically adjusting the detection scope. The algorithm dynamically terminates when no more cycles are found or when k exceeds the graph diameter. This dynamic approach allows the system to adapt message transmission to the actual graph structure and cycle distribution, reducing unnecessary messages compared to a static BFS approach.