GPU inference on databases that exceed GPU memory and RAM using GPU Direct Storage

[4D0R]

Has there been any exploration of using the cufile libraries to perform MSA search on databases directly from disk? With modern NVMe and networking you can get some pretty remarkable throughput and low latency with GPU Direct Storage (GDS), and this seems like a strong fit for the technology. Even if a database doesn't exceed memory, it could be helpful to free up the memory on the system for other tasks.

[fkallen]

In our current version, the database is provided as a memory-mapped file via mmap. This should also allow using databases which exceed host memory.
We have not considered using cuFile.