Abstract
Data-driven modeling of biological systems such as protein-protein interaction networks is data-intensive and combinatorially challenging. Backtracking can constrain a combinatorial search space. Yet, its recursive nature, exacerbated by data-intensity, limits its applicability for large-scale systems. Parallel, scalable, and memory-efficient backtracking is a promising approach. Parallel backtracking suffers from unbalanced loads. Load rebalancing via synchronization and data movement is prohibitively expensive. Balancing these discrepancies, while minimizing end-to-end execution time and memory requirements, is desirable. This paper introduces such a framework. Its scalability and efficiency, demonstrated on the maximal clique enumeration problem, are attributed to the proposed: (a) representation of search tree decomposition to enable parallelization; (b) depth-first parallel search to minimize memory requirement; (c) least stringent synchronization to minimize data movement; and (d) on-demand work stealing with stack splitting to minimize processors’ idle time. The applications of this framework to real biological problems related to bioethanol production are discussed.