Dike: Revisiting Resource Management for Distributed Deep Learning

The recent adoption of deep learning for diverse applications has required scaling infrastructures both horizontally and vertically. As a result, efficient resource management for distributed deep learning (DDL) frameworks is becoming increasingly important. However, existing techniques for scaling DDL applications rely on general-purpose resource managers originally designed for data intensive applications. In contrast, DDL applications present unique challenges for resource management as compared to traditional big data frameworks, such as a different master-slave communication paradigm, deeper ML models that are more computationally and network bound than I/O, and use of heterogeneous resources (GPUs, TPUs, and variable memory). In addition, most DDL frameworks require data scientists to manually configure the task placement and resource assignment to execute DDL models. In this paper, we present Dike, an application scheduler framework that transparently makes scheduling decisions for placement and resource assignment to DDL workers and parameter servers, based on the unique characteristics of the DDL model (number and type of parameters and neural network layers), node heterogeneity (CPU/GPU ratios), and input dataset. We have implemented Dike as a resource manager for DDL jobs in Tensorflow on top of Apache Mesos. We show that Dike significantly outperforms both manual and static assignment of resource offers to Tensorflow tasks, and achieves at least 95% of the optimal throughput for different DDL models such as ResNet, Inception.