This paper presents a hybrid task scheduling framework for cloud-based big data systems that aims at three main objectives: to improve the system's performance, to decrease the expenses, and to increase the energy efficiency. The conceived system combines a rule-based decision engine with a Long Short-Term Memory (LSTM)-based resource prediction model, enabling real-time job assignment based on task urgency, data locality, and system state. The framework is at the top of Apache YARN; thus, it is compatible with batch jobs (via Hadoop/Spark) as well as streaming tasks (via Kafka/Flink). We reproduced the experiments on a 50-node cluster (n2-standard-16 instances, 16 vCPUs, 64 GB RAM), using real workloads of 100 GB–1 TB batch jobs and 1K–5K event/sec streams. Some of the metrics for evaluating the performance of the experiments are job completion time, throughput, cost per TB processed, and energy consumption (Joules/TB). The results indicate a 32–50% improvement in performance, up to 54% savings in cost when using spot instances, and a 25% reduction in energy consumption compared to baseline schedulers such as YARN, Kubernetes, and Spark.
Sultan et al. (Sun,) studied this question.