利用工作分配來提升數據中心網路中服務鏈的部署

Abstract

Middlebox在現今網路中扮演著非常重要的角色，像是防火牆(firewall)、代理伺服器(Proxy)、網路地址轉換(Network Address Translation)等。然而往往我們需要封包依序經過這些middlebox，這樣的需求我們稱為服務鏈。傳統的middlebox大多是高成本並且位置固定的硬體設備，這使得在部署上有所限制，並且導致浪費許多頻寬。網路功能虛擬化能夠將這些middlebox轉換成軟體並且能在一般伺服器上執行，這使得在部署上更有彈性。問題在於如何去部署服務鏈才會更好。由於在資料中心網路或是企業網路中往往有許多的job，我們應該同時分配這些job來做服務鏈的部署才能得到最佳解。並且過去的研究皆並未考慮到使用的伺服器數目，這可能導致電力耗費提高，因此我們尋找出頻寬與運算資源以及使用的伺服器數目之間的關聯性，並且將整個問題公式化成最佳化問題，因為這個最佳化問題十分困難，因此我們提出一個rack aware的演算法有效率地解決此問題。經由模擬實驗的結果證明加入job分配能有效提升服務鏈部署的表現，並且我們提出的演算法能大幅減少運算時間，同時也能有效地找到接近最佳解的解。

Middlebox, such as firewall, proxy and NAT, play an important role in the existing network. However, we usually need traffic go through these middlebox in specific order. We call these requests to Service chain. Most of middleboxes are expensive hardware-based appliances with fixed placement. It is not flexible to deploy and might cause we waste lots of bandwidth. Network function virtualization transforms these middlebox to software to be executed on general-purpose servers. It helps us do deployment more flexible. The question is how to do service chain deployment would be better. Due to there have many jobs in the data center networks and enterprise networks, we should do service chain deployment and job dispatching together so as to find the global optimal solution. Moreover, most prior works do not aware the number of used servers and cause the higher electricity cost. Therefore, we discover the relation among the bandwidth, CPU resource and number of used servers. And we formulate the service chain deployment and job dispatching problem as an integer linear programming model. However, it is very difficult to solve this optimization model. Hence, we propose a Rack-aware Service-chain deployment and Job dispatching (RSJ) algorithm to find the solution effectively and efficiently. The simulation result shows that we can improve the service chain deployment by doing job dispatching together. Our proposed algorithm significantly reduces the processing time and also can find a solution close to the optimal solution.