Add support for GPUDirect Storage (Direct SSD → GPU memory transfers)

[zealbell]

Description

Currently, TornadoVM handles data transfers by staging through host (CPU + system RAM) before moving data to GPU VRAM. While this works, it adds latency and CPU overhead for data-heavy workloads.

NVIDIA provides GPUDirect Storage (GDS), and Microsoft/AMD provide similar technologies (DirectStorage, Smart Access Storage), that enable direct transfers from NVMe SSDs to GPU VRAM, bypassing CPU and system RAM.

This allows:

• Lower CPU utilization (frees cycles for other tasks).
• Higher throughput for large datasets.
• Lower latency for streaming workloads (games, ML, big data).

---

Proposal

Add support for GPUDirect Storage (or an abstraction that uses it where available) so TornadoVM kernels can load data directly from storage into GPU memory.

Possible approaches:

• Expose an API that allows TornadoVM users to request file → device buffer transfers via GPUDirect.
• Under the hood, integrate with NVIDIA cuFile APIs (or equivalent for other backends when/if they exist).
• Fallback gracefully to the current CPU+RAM staging path if GDS is not available.

---

Use Cases

• Machine learning workloads where datasets are too large for CPU RAM but can stream from NVMe.
• Scientific workloads involving massive data sets (climate, genomics, simulations).
• High-performance data processing pipelines.
• Game engines written in Java/JVM languages benefiting from DirectStorage.

---

Questions

1. Is GPUDirect Storage (or similar direct-storage-to-accelerator APIs) on the TornadoVM roadmap?
2. Would the maintainers be open to contributions in this area if a prototype integration is developed?

[kotselidis]

Hallo @zealbell Thank you for your interest in TornadoVM and for your suggestions. We are not too familiar with GDS but it definitely sounds interesting. We will spend sometime to better understand GDS and how it could be implemented in TornadoVM and will come back to you. If you have a proposal on how it could be integrated please let us know.