Traffic engineering in data center networks
Seo Taek Kong · Illinois Digital Environment for Access to Learning and Scholarship (University of Illinois at Urbana-Champaign) · 2017
Commodities in data centers today are often connected in a switch-centric approach to reduce link crosspoints. Today's data centers consist of a large number of servers that need to communicate among another. When data rates approach the capacities of links, congestion may occur or transmission delays increase resulting in a decrease of network throughput. Hence, sophisticated routing schemes are necessary to maximize the throughput of the network. It is possible to compute the optimal routing scheme via linear programming. However, linear programming algorithms have insufficient computation complexity to be useful in large data centers. Proposed in this thesis are two randomized algorithms for load balancing that achieve a suboptimal routing scheme with significantly reduced runtime. The proposed algorithms demonstrate performance very close to the solution obtained by multi-path routing when computed through a convex-programming solver. Further, the runtimes are incomparably faster than the convex-programming solver, even when the solver utilizes multi-threading to compute the optimal solution.