We are using CUDA-aware OpenMPI and seeing large overhead originating from cuPointerGetAttribute and cuMemRetainAllocationHandle. They are calling cudbgMain, cuELGApiInit, and cuVDPAUCtxCreate under the hood. I do not know exactly what those internal APIs are for, but I guess they are related with profiling, logging, and debugging. Is there a public API that enables/disables those? Or are they always called?
