Fregata: A Low-Latency and Resource-Efficient Scheduling for Heterogeneous Jobs in Clouds

An increasing number of large-scale data analytics frameworks move towards larger degrees of parallelism aiming at low-latency guarantees. It is challenging to design a scheduler with low latency and high resource utilization due to task dependency and job heterogeneity. The state-of-the-art schedulers in cloud/datacenters cannot well handle the scheduling of heterogeneous jobs with dependency constraints (e.g., dependency among tasks of a job) for simultaneously achieving low latency and high resource utilization. The key issues lie in the scalability in centralized schedulers, ineffective and inefficient probing and resource sharing in both distributed and hybrid schedulers. To address this challenge, we propose Fregata, a low-latency and resource-efficient scheduling for heterogeneous jobs with constraints (e.g., dependency constraints among tasks of a job) in clouds. Fregata first uses the machine learning algorithm to classify jobs into two categories (high priority jobs and low priority jobs) based on the extracted features. Next, Fregata splits the jobs into tasks and distributes the tasks to the master nodes based on task dependency and the load of master nodes. Then, Fregata utilizes the dependency information of tasks to determine task priority (tasks with more dependent tasks have higher priority), and packs tasks by leveraging the complementary of tasks' requirements on different resource types and task dependency. Finally, the master nodes distribute tasks to workers in the system based on priority of tasks and workers and the resource demands of tasks and the available resources of workers. To test the performance of Fregata, we conduct tracedriven experiments. Extensive experimental results based on a real cluster and Amazon EC2 cloud service show that Fregata achieves low-latency and high resource utilization compared to existing schedulers.