使用軟體定義網路之以網路感知為基礎資料中心虛擬機的負載平衡

Abstract

在現行的雲端數據中心中經由虛擬機遷移方法進行的負載平衡方法仍然存在兩個主要的應用性能問題：產生系統負載平衡度和直到平衡狀態的總時間。多年來，為改進這些指標進行了研究。然而，由於這些問題相互影響，所以同時實現兩個指標的實質性進步是一個挑戰。在這項工作中，我們研究虛擬機器負載平衡的問題，考慮到上述問題對應用程序性能的影響。我們將這個問題視作為多商品最低成本流問題的變化問題，並顯示如何將不平衡程度和總遷移時間同時進行優化。由於這樣的問題是NP-hard，而想在實際數據中心讓應用程式在多項式時間解決是不實際可行的，我們還提出了一種基於蟻群優化算法的啟發式方法。通過模擬結果表示，與最小成本流問題的混合整數規劃公式給出的最優解相比，我們的啟發式給出了良好的性能結果。此外，我們比較我們的方法結果與幾個主要相關作品的結果。由於沒有一個相關的工作直接關注同時考慮到兩個主要指標的負載平衡，所以提供了關於如何比較這些工作的結果的另外的推理。實驗結果表明，與相關工作中提出的現有方法相比，我們的方法平均實現了兩倍快的遷移時間。

Despite being industry standard, load balancing via virtual machine migration methods in modern cloud data centers still exhibit two major performance issues it has on applications: resulting system load balancing degree and total time till balanced state. Over the years research has been conducted to improve these metrics. Yet due to the fact that these issues influence each other, it has been a challenge to achieve substantial advancement in both metrics simultaneously. In this work we study the problem of VM load balancing considering effects mentioned issues have on the applications performance. We formulate this problem as a variation of multi-commodity minimum cost flow problem and show how both resulting degree of imbalance and a total migration time can be optimized together. Because such problem is NP-hard and is not realistic to solve in polynomial time for application in real data centers, we also propose a heuristic method based on Ant Colony Optimization family of algorithms. Through simulation results, we show that our heuristic gives good performance results compared to optimal solutions given by mixed integer programming formulation of minimum cost flow problem. Furthermore, we compare our method results with results from several major related works. Due to the fact that none of the related works directly focus on load balancing with consideration of both main metrics at the same time, additional reasoning on how the results of these works can be compared has been provided. The outcome of experiments show that our method achieves on average twice faster migration time compared to existing methods proposed in related works.