: Scaling the number of I/O threads to maximize bandwidth.

What is your primary storage architecture?

“I created multiple batch_ids and io_batch_params and submitted all of them in a loop ... it seems that multiple batches cannot be submitted at the same time. ... I'm getting approximately 0.7 GB/s for all data sizes.” forums.developer.nvidia.com · 2 years ago Key Considerations for Use

: Recent versions (12.2+) support new memory types for cuFile APIs and non- O_DIRECT file descriptors. Find the right benchmarking setup for you

: It is often used alongside gdscheck to verify that the file system and hardware (such as NVMe drives or RDMA-enabled storage) are correctly configured to support GDS.

The gdsio utility is the standard tool used to evaluate the performance of GDS-enabled systems, offering capabilities similar to the general-purpose fio tool but specifically optimized for GPU memory paths.