A common parallel framework for LLP combinatorial problems
This provides a unified framework for parallelizing combinatorial problems, reducing programmer effort, though it appears incremental as it builds on existing lattice-linear predicate concepts.
The authors tackled the problem of requiring custom lock-free parallel algorithms for combinatorial optimization by proposing a general-purpose runtime, LLP-FW, which solves such problems by advancing forbidden local states in parallel, showing it compares favorably to hand-tuned solutions in most cases.
Traditional lock-free parallel algorithms for combinatorial optimization problems, such as shortest paths, stable matching, and job scheduling require programmers to write problem-specific routines and synchronization code. We propose a general-purpose lock-free runtime, LLP-FW that can solve all combinatorial optimization problems that can be formulated as a Lattice-Linear Predicate by advancing all forbidden local states in parallel until a solution emerges. The only problem-specific code is a definition of the forbiddenness check and a definition of the advancement. We show that LLP-FW can solve several different combinatorial optimization problems, such as Single Source Shortest Paths (SSSP), Breadth-First Search (BFS), Stable Marriage, Job Scheduling, Transitive Closure, Parallel Reduction, and 0-1 Knapsack. We compare LLP-FW against hand-tuned, custom solutions for these seven problems and show that it compares favorably in the majority of cases.