Epic: GPU #7712

Context

So far we've only spent time optimizing the GPU kernels themselves but not how to get the data to the GPU fast. Prior art here is Onur's exploratory work in this area. As a first I'll capture the current GPU data loading state in Vortex which is expected to be slow.

Benchmarking copy modes

Insight => Copy performance does not differ between pageable host memory and pinned host memory. The main difference being that copy calls can return immediately (C style async) in case the memory is pinned to physical addresses. WRITECOMBINED yields significantly slower copy performance for cuMemHostAlloc (see #7815 (comment)).

Benchmark perf between

• copies from host memory to GPU
• NVMe to CPU
• object storage to GPU

GPU memory pooling

• GPU allocation is very expensive => buffer pool

Zero Copy from NVMe and Object Storage

• from NVMe to the GPU
• from object storage to the GPU

cuDF

• cuDF export
• pyvorrtex with cuDF

Vortex Integration

• run the ideal respective modes as part of the scan, bypass host memory if possible
• come up with a decision mechanism if and which parts of the scan should run on the GPU

Spiral Integration

• decompress a Spiral column on the GPU

• bench: CUDA sync load-to-device benchmark #7799
• bench: CUDA host-to-device copy modes #7815