Load Balancing and Capacity Optimization for Network Function Virtualization and Cloud-Edge Computing

• Main Authors: Thai, Minh-Tuan, 蔡明俊
• Other Authors: Lin, Ying-Dar
• Format: Others
• Language: en_US
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/tq29qz

博士 === 國立交通大學 === 電機資訊國際學程 === 106 === Thanks to the concept of network function virtualization (NFV), a composite network service can be provided through a service chain, which is a pre-defined sequence of virtualized network functions (VNFs) being applied to data flows. To accommodate user requests for such a service, providers need to carry out service chaining by selecting appropriate VNFs and network paths to construct service paths through which data flows are steered. Prior service chaining systems address server and network load balancing issues individually, which cannot ensure efficient system performance. To this end, in this dissertation, we first study the design, implementation, complexity analysis, and evaluation of Nearest First and Local-Global Transformation (NF-LGT) algorithm, which jointly supports server and network load balancing for chaining VNFs in data center environment. The algorithm firstly constructs service chains by a greedy strategy, which considers both server and network latency. A searching technique, which replaces a selected VNF with another candidate and swaps the order of VNFs in service chains, is then applied to improve the service chains. We have implemented the algorithm using Software-defined networking (SDN) and OpenFlow concepts. The numerical results indicate that, compared with a sequential approach, NF-LGT increases the system bandwidth utilization by $45\%$. The results also show that it is worth applying the 2$^{nd}$ phase of our algorithm since it considerably enhances the system performance by 20%. After that, we study Hash-based Traffic Steering on Softswitches (HATS) which is a load balancing scheme for chaining VNFs, with the aim of mitigating the control and data plane overheads of existing methods. Our method uses a flow-hashing technique applied to softswitches to carry out server and network load balancing without triggering the controller. By exploiting the advantages of HATS, we then derive two algorithms, HATS with Flowcell-based Multipathing (HATS-Flowcell) and Dynamic Weight Adjustment for HATS (D-HATS), to address hash collision problems which downgrade system performance. The first algorithm divides an elephant flow into various equal-size flowcells, which are distributed over network paths as individual flows. The second algorithm periodically updates the hashing weights of VNFs and network paths according to their current load status. Our implementation demonstrates that HATS can be readily deployed on commodity network hardware. Furthermore, our experimental results show that D-HATS has roughly the same load balancing performance with Least Load First (LLF), a controller-based service chaining algorithm; while significantly reducing the number of flow entries and service chaining time by 54% and 93%. In addition to the load-balancing solutions for VNF chaining, this dissertation proposes a generic architecture of cloud-edge computing with the aim of providing both vertical and horizontal offloading between service nodes. To investigate the effectiveness of the design in different operational scenarios, we formulate it to a workload and capacity optimization problem with the objective of minimizing the system computation and communication cost. Because such a mixed integer nonlinear programming (MINLP) problem is NP-hard, we further develop an approximation algorithm which applies a branch-and-bound method to obtain optimal solutions iteratively. Experimental results show that such cloud-edge computing architecture can significantly reduce total system costs by about 34%, compared to traditional designs which only support vertical offloading.