資料中心的平行分級流量排程

Abstract

資料中心的流量排程問題需要低延遲和分級處理。我們提出動態分級類神經網路(GRNN)，與現有的類神經網路不同的地方是，GRNN可以完全離散、平行運算，接著我們使用GRNN來實現資料中心的排程機，GRNN可以達到接近最佳化的效果，運算的時間複雜度上限是O(N)，最後，透過模擬可以發現，實際的計算時間與上限相比是接近常數。

Flow scheduling in datacenter requires extremely low delay and prioritized processing. In this paper, we propose Grouping Ranking Neural Network (GRNN) paradigm. Unlike the existing recurrent neural network models, GRNN paradigm is capable of operating in f fully parallel discrete-time manner. We then proposed GRNN-based scheduler for datacenter. The flow scheduling in datacenter is a kind of joint competition group k-winner-take-all problem. GRNN paradigm could achieve near-optimal performance. And the upper bound of time complexity is O(N), where N is the problem size. Via the simulation, we demonstrate that the convergence time of GRNN scheduler is near constant, compared with upper bound.