[Question] Identifying All Instances of cudaMemcpy with Pageable Host Memory

[user157267]

This question isn’t strictly confined to the nsys profiler, though nsys serves as my starting point.

Here’s the motivation behind this query:
When using the nsys profiler, I detected certain cudaMemcpy operations involving pageable host memory. However, given the sampling-based nature of profiling, I’m concerned that nsys may not capture every single cudaMemcpy operation on pageable memory.

I’m interested in determining if there’s a reliable method to show all cudaMemcpy operations that utilize pageable memory, whether through nsys or other profiling tools.
The rationale behind this is that in my multi-streaming context, pageable host memory is highly unfavorable. This is because copies from&to pageable memory will cause synchronization issues.

Thank you!

[hwilper]

Nsight Systems has both API trace and sampling functionality. And in the case of CUDA, all of the functions are traced, rather than being sampled.

We get a callback every time the function is called, so you are seeing every instance.

I would also like to suggest that you consider running the expert systems rule here:

[user157267]

Thank you. I explored the Expert Systems , and it provided a useful summary regarding Pageable Memcpy, along with information on other aspects such as sync Memcpy.

I also learned that all relevant function calls are listed there. With this assistance, I’ve managed to eliminate all Pageable Memcpy operations from my program. Thanks!