Performance Evaluation of the Weighted Least Connection Scheduling for Datacenters with BigHouse Simulator

In this paper, we investigate the performance of a weighted least connection algorithm for scheduling jobs in datacenters. The novelty of the proposed algorithm is that the weights for the compute units in the datacenter are determined based on their current dynamic power consumption. The algorithm is implemented inside the BigHouse simulation framework and compared against the default least utilized scheduling approach of the framework. Simulation experiments show that the proposed algorithm provides significantly better performance for large number of large queries per second (QPS) values as well as lower power consumption. In addition, the computational runtime is linear with respect to the increase in queries per second. However, while linear, the computational runtime is longer than that of the default scheduler due to the increased computational complexity required to determine where a job should be placed. These results indicate a tradeoff between performance (i.e., latency of all scheduled jobs) and computational runtime of the scheduling algorithm.