Cross-Architecture Performance Prediction (XAPP):

mitsue-stanley
Published  . 0 views
↓ Download
Cross-Architecture Performance Prediction (XAPP):
1 / 1
Cross-Architecture Performance Prediction (XAPP): - slide 1 of 19 Cross-Architecture Performance Prediction (XAPP): - slide 2 of 19 Cross-Architecture Performance Prediction (XAPP): - slide 3 of 19 Cross-Architecture Performance Prediction (XAPP): - slide 4 of 19 Cross-Architecture Performance Prediction (XAPP): - slide 5 of 19 Cross-Architecture Performance Prediction (XAPP): - slide 6 of 19 Cross-Architecture Performance Prediction (XAPP): - slide 7 of 19 Cross-Architecture Performance Prediction (XAPP): - slide 8 of 19 Cross-Architecture Performance Prediction (XAPP): - slide 9 of 19 Cross-Architecture Performance Prediction (XAPP): - slide 10 of 19 Cross-Architecture Performance Prediction (XAPP): - slide 11 of 19 Cross-Architecture Performance Prediction (XAPP): - slide 12 of 19 Cross-Architecture Performance Prediction (XAPP): - slide 13 of 19 Cross-Architecture Performance Prediction (XAPP): - slide 14 of 19 Cross-Architecture Performance Prediction (XAPP): - slide 15 of 19 Cross-Architecture Performance Prediction (XAPP): - slide 16 of 19 Cross-Architecture Performance Prediction (XAPP): - slide 17 of 19 Cross-Architecture Performance Prediction (XAPP): - slide 18 of 19 Cross-Architecture Performance Prediction (XAPP): - slide 19 of 19
Description: Cross-Architecture Performance Prediction (XAPP): Using CPU to predict GPU Performance Newsha Ardalani Clint Lestourgeon Karthikeyan Sankaralingam Xiaojin Zhu University of Wisconsin-Madison Executive Summary Problem: GPU programming is

Related Topics

Download Presentation

"Cross-Architecture Performance Prediction (XAPP):" is the property of its rightful owner. Permission is granted to download and print the materials on this website for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.

Presentation Transcript

slide1. Cross-Architecture Performance Prediction (XAPP): Using CPU to predict GPU Performance Newsha Ardalani
Clint Lestourgeon
Karthikeyan Sankaralingam
Xiaojin Zhu University of Wisconsin-Madison<br>
slide2. Executive Summary Problem: GPU programming is challenging
Not clear how much speedup is achievable
Goal: Save programmers’ time from unnecessary porting effort
Insight:
Speedup is correlated with program properties and hardware characteristics.
Machine learning can learn the correlation
Results:
27% relative error across 24 benchmarks 2<br>
slide3. Outline Problem Statement
Insight
Overviews
Machine Learning Technique
Results
Future work 3<br>
slide4. Problem Statement For a given GPU platform, predict optimized GPU execution time obtainable for any CPU application prior to starting the code development process 4<br>
slide5. Insight GPU performance GPU hardware characteristics
Number of cores
GPU Performance Inherent program properties
Available parallelism: Inherent program property
Cache miss: platform-dependent program property
GPU Execution time = 5<br>
slide6. 6 GPU Execution Time … … Feature Vector Fx ( feature vector) Predicted
GPU Execution Time Overview Dynamic Binary Instrumentation Dynamic Binary Instrumentation<br>
slide7. Machine Learning Technique: 7 … … Step-wise Regression Step-wise Regression Step-wise Regression Majority Selection<br>
slide8. Outline Problem Statement
Insight
Overviews
Machine Learning Technique
Results
Future work 8<br>
slide9. Results 9<br>
slide10. Experimental Setup Accuracy results on GPU GTX 750
Training Set: 112 datapoints
Rodinia, Lonestar, Parboil, Parsec Subset, NAS subset
Test Set: 24 datapoints
Feature Vector: 31 program properties
Program properties collected using MICA and Pin
Execution time measured on real GPU hardware 10<br>
slide11. Accuracy results 11 Accuracy on platform 1 (GTX 750): 27% relative error
Accuracy on platform 2 (GTX 660): 36% relative error<br>
slide12. 12 GPU Execution Time … … Feature Vector Fx ( feature vector) How to use our tool? Model Construction Phase<br>
slide13. How to use our tool? Usage Phase 13 Speedup Prediction<br>
slide14. 14 GPU Execution Time … … Feature Vector Fx ( feature vector) One-time Cost Overhead<br>
slide15. 15 GPU Execution Time … … Feature Vector Fx ( feature vector) Predicted
GPU Execution Time Recurring Overhead<br>
slide16. 16 … … Fx ( feature vector) What it is not?<br>
slide17. Also in the paper Why Ensemble Prediction?
Model Interpretation
End-to-End Case Studies 17<br>
slide18. Summary 18 FPGA XeonPHI<br>
slide19. Questions? 19<br>