Presentation
PoDD: Power-Capping Dependent Distributed Applications
SessionPower and Scale
Event Type
Paper
TP
Deep Learning
Energy
Extreme Scale Computing
Machine Learning
Performance
Power
Resource Management
Runtime Systems
Simulation
BP Finalist
BSP Finalist
TimeTuesday, 19 November 20194pm - 4:30pm
Location405-406-407
DescriptionPower budgeting (or capping) has become essential for large-scale computing installations. As these systems scale out, they are able to concurrently execute dependent applications that were serially processed previously. This reduces resource usage and execution time as they communicate at runtime instead of through disk. One challenge for power budgeting systems is how to power cap dependent applications for high performance. Existing approaches, however, have major limitations: (1) poor practicality, due to dependence on application profiles, (2) only optimize for power reallocation between nodes, without considering node-level optimization.
We propose PoDD -- a hierarchical, distributed, dynamic power management system for dependent applications. We implement it on a 49-node cluster and compare it to SLURM, a state-of-the-art power management system, but not considering coupling, and PowerShift, a power capping system for dependent applications without node-level optimization. On average, PoDD increases performance over SLURM by 14-22%, over PowerShift by 11-13%.
We propose PoDD -- a hierarchical, distributed, dynamic power management system for dependent applications. We implement it on a 49-node cluster and compare it to SLURM, a state-of-the-art power management system, but not considering coupling, and PowerShift, a power capping system for dependent applications without node-level optimization. On average, PoDD increases performance over SLURM by 14-22%, over PowerShift by 11-13%.
Download PDF
Archive