分析了并行优化算法中同步运算与异步运算的优缺点。提出完全异步的PGD算法。并且在一定的条件下。给出了算法的收敛性证明。最后结合大规模分布式并行计算机系统曙光-2000做出数值试验。结果说明异步的并行优化算法的效率高于同步的算法。
The paper analyzes the advantages and disadvantages of synchronous and asynchronous operations in parallel optimization algorithm, and presents the complete asynchronous PGD algorithm and under a certain condition, the proof of its global convergence is given. Finally, numerical tests are provided combined with the distributed large-scale parallel computer system, Dawning 2000, and its results show that the efficiency of asynchronous parallel optimization algorithm is better than that of its synchronous counterpart.