HPC Workload Fingerprinting for Tunable Hardware Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
High-performance computing (HPC) systems face challenges in balancing power consumption and performance, often requiring repetitive optimization processes that are inefficient and node-specific, and struggle with optimizing multiple hardware parameters simultaneously.
Innovation Solution
An unsupervised machine learning technique is used to classify workloads into classes based on performance profiling counters, determining optimal settings for tunable hardware parameters like CPU and GPU frequencies, which are then stored in a knowledge database for efficient application across all nodes, reducing computational overhead and improving system performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional optimization methods are used for each node individually, then node-specific performance can be optimized, but the optimization process becomes repetitive and computationally expensive across the entire HPC system
Solution Approach 1:
The patent segments the HPC system into workload categories (e.g., CPU-bound, GPU-bound, mixed-workload) and creates separate optimization profiles for each category. This allows the system to apply pre-determined optimal settings based on workload type rather than performing individual optimization for each node, reducing repetitive computation while maintaining precision.
Solution Approach 2:
The patent performs optimization analysis in advance during system setup or idle periods, determining optimal hardware parameter settings for different workload categories before actual computation begins. These pre-computed optimization profiles are then applied automatically during runtime, eliminating the need for repetitive optimization during productive computation time.
2Productivity
If multiple hardware parameters are optimized simultaneously, then system performance can be improved, but the complexity of the optimization process increases significantly
Solution Approach 1:
The patent decomposes the complex multi-parameter optimization problem into separate optimization tasks for different hardware components (CPU, GPU, memory, interconnect) and workload types. Each component can be optimized independently with its own set of parameters, reducing the overall complexity while maintaining the ability to tune multiple parameters across the system.
Solution Approach 2:
The patent focuses on changing key hardware parameters (such as frequency scaling, voltage settings, memory allocation) that have the most significant impact on performance. By identifying and optimizing only the most critical parameters rather than all possible parameters, the system achieves performance improvement without excessive complexity.
3Use of energy by moving object
If power consumption is reduced to balance energy costs, then operating costs decrease, but system performance and job completion times increase
Solution Approach 1:
The patent implements dynamic power management that adjusts hardware parameters in real-time based on actual workload characteristics and system conditions. Rather than using static power caps that uniformly limit performance, the system dynamically scales power consumption to match actual computational needs, reducing waste during low-utilization periods while maintaining high performance during intensive workloads.
Solution Approach 2:
The patent changes hardware parameters such as CPU/GPU frequency and voltage dynamically based on workload requirements. By adjusting these parameters to match actual computational demands rather than maintaining constant high settings, the system reduces power consumption during lighter tasks while preserving the ability to deliver peak performance when needed.
Data Source
AI summary
A method for optimizing operations of high-performance computing (HPC) systems includes collecting data associated with a plurality of workload performance profiling counters associated with a workload during runtime of the workload in an HPC system. Based on the collected data, the method includes using a machine-learning technique to classify the workload by determining a workload-specific fingerprint for the workload. The method includes identifying an optimization metric to optimize during running of the workload in the HPC system. The method includes determining an optimal setting for a plurality of tunable hardware execution parameters as measured against the optimization metric by varying at least a portion of the plurality of tunable hardware execution parameters. The method includes storing the workload-specific fingerprint, the optimization metric, and the optimal setting for the plurality of tunable hardware execution parameters as measured against the optimization metric in an architecture-specific knowledge database.