Repository navigation
Need DMA/RDMA support #482
Description
Activity
Have you looked at how libFabric may help here? It does some work with Level Zero under the hood if you have a supported NIC / Fabric interface it may have your answer.
RDMA is a multiple driver support problem (i.e. network + storage OR network + gpu/offload device)
<< Level Zero Loader is not a driver at all, just an interface to one type i.e. offload devices >>For offload, DMA is added by the individual drivers, not the L0 Loader which is vendor and HW agnostic. RMDA ties in by allowing another driver the NIC, to directly interface with exported memory regions from the GPU/NPU. This repository is simply the loading layer to individual drivers which choose to support the L0 Interfaces. i.e. at current Intel GPU's and Intel NPU's are directly authoring drivers for this, they both have public repos which can be commented in for issues like this and pending on your HW it may already be supported and they both already support DMA. Unsure on the NPU side, but the GPU driver definately supports exporting the memory region for apps or other drivers to use.
So in summary DMA is probably already enabled, and it's chosen based on buffer locality and the operation being invoked at the time of invocation. RDMA is tying together NIC + Level Zero drivers, which is something libfabric is probably done or working on.
I am not aware of a storage driver, but I don't play in that space. I suspect that some forms of storage drivers may already be using libfabric which would then potentially inherit the benefits.
Closing this, unless there is something more specific to address.
As I want to see Intel grow, I want Intel to adapt DMA/RDMA as well. This will help quite a lot, if we have less RAM, and a minimal CPU. We can directly load from NVMe to GPU and write back as well.
NVIDIA has cuFille, GDS. AMD has hipFile. But when I searched for Intel, I found nothing.
NB: Willing to work on this one. Cause I want to grow as well.
Ref: