Since correlationId is now a uint32_t, if CUPTI is enabled for an extended period (e.g., hours or days) and the ID space is exhausted, how is correlationId expected to behave? Does it wrap around, reset, or follow some other convention?
Moreover, I wonder if there’s any known memory leak problem if enabling CUPTI for a long time.
Hi Keren,
Regarding the correlation ID: if the value exceeds the 32-bit range, it will wrap around. In a future release, we are considering using unsigned 64-bit for correlation id
On the memory leak question: CUPTI attempts to free all memory that is no longer needed. If you are observing leaks, could you share more details about them? In particular, are these on the GPU side or CPU side?
For CPU memory, there are certain allocations that are intentionally left for the user to free. The cases where the user is responsible are documented here: https://docs.nvidia.com/cupti/main/main.html#memory-overhead
Please let us know what type of leak you’re seeing (GPU vs CPU) and, if possible, provide a small reproducer or more specifics so we can help investigate.
Got you. We haven’t observed memory leaks yet, but just wanted to confirm. Thanks for the answer!
This topic was automatically closed 5 days after the last reply. New replies are no longer allowed.