[SYSTEMDS-3024] Improve performance by batching data descriptor transfers

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: SystemDS 2.1
• Fix Version/s: SystemDS 3.0
• Component/s: NativeRuntime
• Labels: CUDA GPU codegen
• Epic Link: spoof-cuda
• Assignee: Mark Dokter
• Reporter: Mark Dokter
• Created: 14/Jun/21 23:28
• Updated: 25/Mar/22 00:32
• Resolved: 25/Mar/22 00:32

Description

The spoof cuda operators do several little cudaMemcpy() invocations per operator execution. By transferring all data in one go the overhead can be reduced. In addition, using asynchronous copies can further improve things and are a first step towards using more asynchronicity in the GPU operations.

Issue Links

• links to: GitHub Pull Request #1567