On the Stability of Server Selection Algorithms Against Network Fluctuations

When a set of servers are available for a certain client-server style service, a client selects one of the servers using some server selection algorithm. The best-server algorithm in which a client selects the best one among the available servers by some metric is widely used for performance. However, when a network fluctuation occurs, the best-server algorithm often causes a sudden shift of the server load and could amplify the fluctuation. Reciprocal algorithms in which a client selects a server with a probability reciprocal to some metric are more stable than the best-server algorithm in the face of network fluctuations but their performance is not satisfactory. In order to investigate trade-offs between the stability and the performance in server selection algorithms, we evaluate the existing algorithms by simulation and visualize the results to capture the stability of the server load. From the simulation results, we found that the performance problem of the reciprocal algorithms lies in selecting high-cost servers with a non-negligible probability. Therefore, we propose a 2-step server selection scheme in which a client selects a working-set out of available servers for efficiency, and then, probabilistically selects one in the working-set for resiliency. We evaluate the proposed algorithm through simulation and show that our method is adaptive to environments, easy to load-balance, scalable, and efficient.