GPU Application Performance Estimation via CPU-GPU Delay Simulation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current methods fail to accurately estimate the performance of Graphical Processing Unit (GPU) applications on new computing machines with different GPU-link performance ratios, especially when the internal structure of the target GPU application is unknown, due to varying CPU-GPU communication and computation dependencies.

Innovation Solution

A method is introduced to simulate a delayed-communication environment on a current computing machine by adding a delay to CPU-GPU communication, executing the target GPU application, measuring its performance, and estimating the performance of the new machine based on these measurements.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If GPU computation speed is increased, then computational performance is improved, but the accuracy of performance estimation for new machines deteriorates due to varying CPU-GPU communication performance ratios

Engineering Contradiction:
ImproveGPU computational performanceVSAvoidperformance estimation accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent changes the communication performance ratio parameter between CPU and GPU by introducing artificial delays. By adjusting the delay parameter to match different GPU-to-CPU performance ratios (e.g., 1.5x, 2x, 5x), the system can accurately estimate performance on future machines with different hardware configurations without actually having those machines, thus resolving the contradiction between improved GPU performance and maintained estimation accuracy

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent creates a virtual copy of the target machine's communication characteristics by simulating the GPU-link performance ratio on the current machine. Instead of needing the actual new machine hardware, the system copies the essential performance ratio characteristic through software-based delay insertion, enabling accurate performance estimation while maintaining productivity benefits

Inventive Principle:
Principle #26Copying

2Measurement precision

If the internal structure of the target GPU application is analyzed to improve performance estimation accuracy, then measurement precision is improved, but device complexity increases

Engineering Contradiction:
Improveperformance estimation accuracyVSAvoidapplication structure analysis complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent extracts only the essential performance ratio characteristic (GPU-to-CPU communication ratio) from the complex application structure, ignoring other internal details. By taking out just this critical parameter and simulating it through delay insertion, the system achieves accurate performance estimation without the complexity of analyzing the complete application internal structure

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent segments the performance estimation problem into independent components: GPU computation performance and CPU-GPU communication performance. By separating these and independently adjusting the communication ratio through delay parameters, the system simplifies the estimation process while maintaining accuracy, avoiding the need to analyze the entire application structure as a complex whole

Inventive Principle:
Principle #1Segmentation

Data Source

PatentUS20190325549A1Estimating performance of GPU application for different GPU-link performance ratio
Publication Date: 2019.10.24 INTERNATIONAL BUSINESS MACHINE CORPORATION
  • US20190325549A1 patent drawing
  • US20190325549A1 patent drawing
  • US20190325549A1 patent drawing

AI summary

A computer-implemented method is provided for estimating the performance of a GPU application on a new computing machine having an increased GPU-link performance ratio relative to a current computing machine having a current GPU-link performance ratio. The method includes adding a delay to CPU-GPU communication on the current computing machine to simulate a delayed-communication environment on the current computing machine. The method further includes executing the target GPU application in the delayed-communication environment. The method also includes measuring the performance of the target GPU application in the delayed-communication environment. The method additionally includes estimating the performance of the new computing machine having the increased higher GPU-link performance ratio, based on the measured performance of the target GPU application in the delayed-communication environment.