An experimental study of k-splittable scheduling for DNS-based traffic allocation

The Internet domain name system (DNS) uses rotation of address lists to perform load distribution among replicated servers. We model this kind of load balancing mechanism in form of a set of request streams with different rates that should be mapped to a set of servers. Rotating a list of length k corresponds to breaking streams into k equally sized pieces. We compare this and other strategies of how to break the streams into a bounded number of pieces and how to map these pieces to the servers. One of the strategies that we study computes an optimal k-splittable allocation using a scheduling algorithm that breaks streams into at most k ≥ 2 pieces of possibly different size and maps these pieces to the servers in such a way that the maximum load over all servers is minimized. Our experimental study is done using the network simulator SSFNet. We study the average and maximum delay experienced by HTTP requests for various traffic allocation policies and traffic patterns. Our simulations show that splitting data streams can reduce the maximum as well as the average latency of HTTP requests significantly. This improvement can be observed even if streams are simply broken into k equally sized pieces that are mapped randomly to the servers. Using allocations computed by machine scheduling algorithms, we observe further significant improvements.