With ever increasing data volumes, large compute clusters that process data in a distributed manner have become prevalent in industry. For distributed stream processing platforms (such as Storm) the question of how to distribute workload to available machines, has important implications for the overall performance of the system.
We present a workload scheduling strategy that is based on a graph partitioning algorithm. The scheduler is application agnostic: it collects the communication behavior of running applications and creates the schedules by partitioning the resulting communication graph using the METIS graph partitioning software. As we build upon graph partitioning algorithms that have been shown to scale to very large graphs, our approach can cope with topologies with millions of tasks. While the experiments in this paper assume static data loads, our approach could also be used in a dynamic setting.
We implemented our proposed algorithm for the Storm stream processing system and evaluated it on a commodity cluster with up to 80 machines. The evaluation was conducted on four different use cases – three using synthetic data loads and one application that processes real data.
We compared our algorithm against two state-of-the-art scheduler implementations and show that our approach offers significant improvements in terms of resource utilization, enabling higher throughput at reduced network loads. We show that these improvements can be achieved while maintaining a balanced workload in terms of CPU usage and bandwidth consumption across the cluster. We also found that the performance advantage increases with message size, providing an important insight for stream-processing approaches based on micro-batching.