A fast internet latency matrix completion algorithm with slow-start strategy

Latency matrix completion is an important foundation for the optimizations of latency-sensitive applications. But in full decentralized environments, the latency matrix is always incomplete. Owing to the low-rank property of this matrix, this paper introduces a slow-start strategy and then proposes a novel algorithm to complete the matrix approximately with low computation cost. At first we estimate the rank of the latency matrix approximately, and then transform the problem into a couple of convex function minimization problems to solve. By introducing a kind of sub-gradient descending algorithm, the optimized value can be obtained. To improve the convergence speed, we assign the iterative step size by slow-start strategy. Our experiments show that our algorithm can improve the computing speed significantly without losing any estimation accuracy.