MF-Storm: a maximum flow-based job scheduler for stream processing engines on computational clusters to increase throughput

Background A scheduling algorithm tries to schedule multiple computational tasks on a cluster of multiple computing nodes to maximize throughput with optimal utilization of computational and communicational resources. A Stream Processing Engine (SPE) is deployed to run streaming applications (computational tasks) on a computational cluster which helps execution and coordination of these applications. It is observed that there is a gap in the optimal mapping of a computational and communicational load of a streaming application on the underlying computational and communication power of the resources (cluster). Frequently communicated tasks are scheduled at different processing nodes with relatively slow communicating links. This increases network latency with a decrease in resource utilization. Hence, reduces the achieved throughput of the cluster significantly. Methods MF-Storm, a max-flow min-cut based job scheduler is presented to achieve a near-optimum schedule to maximize throughput. It schedules a streaming application by considering the processing, communication demands, available computational and communicational resources in a heterogeneous cluster, dynamically with minimized scheduling cost. To keep the scheduling cost minimum, the scheduler is built in a pipeline with two major stages: in the first stage, the application’s tasks graph is partitioned using the max-flow min-cut algorithm to minimize inter-partition traffic, and in the second stage, these partitions are assigned to computing nodes according to the computational power of the cluster’s nodes. Results Extensive experiments were done to evaluate the performance of MF-Storm using different topologies with multiple scenarios on a physical cluster implementation. Results showed on average 148% improvement in throughput with 30% less computational resources as compared to different state-of-the-art schedulers.

Abstract licence: CC BY

A person counts once a day per paper. We keep no cookie and no address: a reader is a hash that changes every day. Any month under five reads shows no number.