LinkedIn's DuaLip-GPU: PyTorch powers 10x faster extreme-scale optimization

LinkedIn's DuaLip was a distributed solver for linear programs (LPs) that optimized decisions like job matching, metric balancing, and email volume—problems involving hundreds of millions of users and trillions of decision variables. The original implementation ran on a CPU-bound Scala/Spark stack, limiting hardware utilization and making it difficult to extend to new problem formulations. To overcome these bottlenecks, LinkedIn's team rebuilt the core execution engine using PyTorch, leveraging its GPU acceleration and flexible define-by-run paradigm. The result, DuaLip-GPU, replaces expensive matrix factorizations with first-order methods that rely on sparse matrix-vector ops and blockwise projections, all optimized for multi-GPU clusters.

Initial benchmarks show order-of-magnitude speedups compared to the CPU baseline, with efficient scaling across multiple GPUs and a 50% reduction in engineering overhead for new constraints. By exposing the hot path as an explicit dataflow over PyTorch tensors, the system retains the mathematical guarantees of linear programming while gaining the performance and flexibility of modern deep learning frameworks. This approach demonstrates that PyTorch can serve as a powerful engine for non-deep-learning optimization at web scale, potentially influencing how other platforms tackle similar extreme-scale decision problems.