Why cudaStreamCreateWithFlags and cudaFree are very time-consuming?

Accelerated Computing CUDA CUDA Programming and Performance

86108429 April 16, 2019, 9:47am 1

Hello, I use nvprof to profiling my executable program, but I find the cudaStreamCreateWithFlags and cudaFree are very time-consuming:

API calls:   
58.56%  273.858s        16  17.1161s  16.875us  136.929s  cudaStreamCreateWithFlags
33.46%  156.448s        17  9.20280s     924ns  74.3984s  cudaFree
...

Is this phenomenon is normal?

Topic		Replies	Views
Something make me fell confuse when I use the nvprof Visual Profiler and nvprof cuda	0	725	June 4, 2022
cudaFree takes approx 99.5% of total time. CUDA Programming and Performance	2	1605	April 11, 2018
Time of API calls in nvprof's output is consumed in GPU or CPU Jetson TX2	2	565	October 18, 2021
Profiling CUDA Programming and Performance	0	489	August 13, 2015
Profiling CUDA Programming and Performance	2	826	August 17, 2015
Multi-process test cudastreamcreate overhead too long CUDA Programming and Performance	0	638	July 11, 2017
CUDA stream not fully utilize CUDA Programming and Performance	0	454	July 27, 2017
Abnormal overhead during stream creation and data copy. Jetson TX2	4	713	July 28, 2017
cudaMemcpy - consuming lot of execution time CUDA Programming and Performance	0	405	April 16, 2018
What is the defferent between"GPU activities" and "API calls"? Legacy PGI Compilers	3	3192	June 4, 2019

Why cudaStreamCreateWithFlags and cudaFree are very time-consuming?

Related topics