IEEE/ACM International Symposium on Cluster, Cloud and Internet Computing 

Paper Accepted at an Outstanding International Conference

(First author: Kim Yu-chan; corresponding author: Kim Young-jae)

▲ (From top left) Kim Yu-chan, master's student; Kim Ki-hyun, combined master's–doctoral student

(From bottom left) Kim Jin-woo, combined master's–doctoral student; Professor Kim Young-jae

The paper 'DeepVM: Integrating Spot and On-Demand VMs for Cost-Efficient Deep Learning Clusters in the Cloud', written by Kim Yu-chan of the Data-Centric Computing and Systems Laboratory (first author), Kim Ki-hyun (combined master's–doctoral, supervisor Kim Young-jae) and Kim Jin-woo (combined master's–doctoral) with Professor Kim Young-jae (corresponding author), has been accepted for publication at the IEEE/ACM International Symposium on Cluster, Cloud and Internet Computing (IEEE/ACM CCGRID).

The paper introduces DeepVM for deep learning workloads on cloud-based GPU clusters, proposing a way to maximize cost efficiency by combining spot VMs and on-demand VMs. DeepVM's cost savings and efficiency gains were demonstrated through evaluation in a real AWS environment. The research makes deep learning accessible to a wider range of users and optimizes GPU resource utilization, enabling more effective and economical deep learning training.

Kim Yu-chan, the paper's first author, said: "I immersed myself deeply in this topic during my first year of graduate school, and I am delighted with an achievement that came after long effort. There were difficulties along the way, but the help and encouragement of those around me kept me going. I am grateful to Professor Kim Young-jae for his guidance, and I hope graduate students will stay steady to the end, develop their own research and achieve good results."

▲ Overview of DeepVM

▲ Cost-efficiency evaluation of cluster configurations using DeepVM

IEEE/ACM CCGRID (IEEE/ACM International Symposium on Cluster, Cloud and Internet Computing) is an internationally recognized conference in cluster, cloud and grid computing. It is listed among the outstanding conferences of the Korean Institute of Information Scientists and Engineers and covers topics such as computing clusters and distributed clouds for AI applications.

References:

Conference: https://2024.ccgrid-conference.org/

Paper: https://arxiv.org/pdf/2403.05861.pdf