DYNAMIC RESOURCE ALLOCATION FOR DEEP LEARNING TRAINING USING TENSORFLOW ON KUBERNETES CLUSTER

• Main Author: Yesa Surya, Rahmad
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/39082
• Institution: Institut Teknologi Bandung

Distributed deep learning training nowadays use static resource allocation. Using parameter server architecture, deep learning training is carried out by several parameter server (ps) nodes and worker nodes. Their numbers are constant while the training is running, hence static. Consider a training running on general purpose cluster where there are possibly some other processess running other than the training. In the middle of its run, some other processess may end, giving free resources on the cluster. The training with static configuration, therefore, cannot utilize those free resources to speed up its process. Besides that, if those free resources are unused for a long time, cluster’s resource become underutilized. In this final project, dynamic resource allocation is designed and implemented for TensorFlow training job in Kubernetes cluster. The implementation is done by creating a component called ConfigurationManager (CM). The role of this component is to know the cluster’s resource information as well as adding more ps and worker node to the training once free resources exist. Training is designed to communicate periodically with CM to report its progress. The experiment shows that the training with dynamic resource allocation has better performance than one with static resource allocation on following metrics: resource usage, epoch time, and total training time. However, training with dynamic resource allocation has slightly lower accuracy than one with static resource allocation.