HPC Node Power Capping by Job Type to Preserve Throughput
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing power capping mechanisms in high-performance computers (HPCs) significantly degrade performance, particularly for compute-bound jobs, and fail to coordinate power capping adjustments across different jobs running on a managed partition.
Innovation Solution
A mechanism that assigns power caps based on the type of job (memory-bound, compute-bound, or mixed-type) to optimize power consumption, allowing for dynamic adjustment of existing power caps to maintain performance and comply with a predefined threshold.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by stationary object
If power capping is enforced on HPC nodes, then power consumption is limited, but job throughput and performance are significantly degraded
Solution Approach 1:
The patent implements dynamic power capping that adjusts power limits in real-time based on job characteristics, workload conditions, and node utilization. Instead of static power caps, the system continuously adapts power allocation to maintain optimal performance while respecting power budgets, resolving the contradiction between power limitation and performance maintenance
Solution Approach 2:
The patent applies differentiated power capping strategies to different nodes within the same partition based on their specific job types and workload characteristics. Compute-bound jobs receive different power allocation than memory-bound jobs, allowing each node to operate at its optimal power-performance point rather than applying a uniform cap across all nodes
2Device complexity
If uniform power capping is applied across all nodes, then power management is simplified, but performance degradation occurs particularly for compute-bound jobs
Solution Approach 1:
The system implements node-specific power capping policies that tailor power limits to the actual workload characteristics of each node. Compute-bound nodes receive higher power allocation during compute-intensive phases, while memory-bound nodes receive appropriate power for memory operations, eliminating the one-size-fits-all approach that degrades performance
Solution Approach 2:
The patent pre-characterizes job types (compute-bound, memory-bound, mixed) before job execution and pre-configures appropriate power capping parameters for each job category. This preliminary classification allows the system to apply optimized power settings from the start of job execution rather than reacting to performance degradation after the fact
3Use of energy by stationary object
If power cap values are reduced to meet power budget, then power consumption compliance is achieved, but performance degradation increases
Solution Approach 1:
The system dynamically changes power cap parameters based on real-time monitoring of both power consumption and performance metrics. When performance degradation is detected, the system adjusts power parameters to restore performance while maintaining overall budget compliance, rather than simply reducing power caps to meet budget constraints
Data Source
AI summary
The invention relates to a computer implemented method for capping the power consumption of a high-performance computer including a plurality of nodes. The plurality of nodes include a first group of nodes of which each node is allocated to a first job, the power capping of each node of the first group of nodes being enforced to a first capping value being dependent on the type of the first job.