simpleStreams: Asynchronous Data Transfer in NVIDIAs CUDA

1
simpleStreamsAsynchronous Data Transfer in
NVIDIAs CUDA

• Supada Laosooksathit

2
Content

• Introduction
• Streams Experiment
• Performance Model
• Conclusion

3
Introduction
4
Huge Data Transfer
5
Requirements
Latency Hiding Strategy
6
Requirements
Streams
7
Streams Experiment

• simpleStreams on CUDA
• Initialize an array to a specific value
• Copy data back to host asynchronously
• Variables
• Number of streams
• Array size

8
Experiment Results
9
Performance Model

• Assumptions
• Kernel execution times are the same
• Memory copy time are the same
• Factors
• Number of Streams

10
Time Diagram Synchronous
11
Time Diagram H2D Asynchronous
12
Time Diagram D2H Asynchronous
13
Time Diagram Both-side Asynchronous
14
Conclusion

• Number of data
• How often the data to be transfer
• Kernel execution time

15
Questions?
16
(No Transcript)