Among cudaEventRecord and nvtxRangePush, whose overhead is smaller? about nvtx

[harrism]

You may find these CUDA event benchmarks useful: https://github.com/harrism/cuda_event_benchmark. In my tests on V100 / Ubuntu 18.04 / CUDA 10.2 / AMD Ryzen 7 3700, a default event record (timing enable) has a throughput of about 400K records per second. With timing disabled it is 10x faster, but if you are intending to use events for timing, go with the lower throughput.

The event approach works well with tools like Google Benchmark, however you may need to take extra steps to flush the L2 cache between kernels if you need to benchmark the performance assuming a cold cache. You can see how we do this in RAPIDS libcudf benchmarks with this class, and an example benchmark that uses it:
https://github.com/rapidsai/cudf/blob/f78f80e94c74c08fface696cfd7e03881b9b0380/cpp/benchmarks/transpose/transpose_benchmark.cu#L46-L49

Note that using CUDA events for timing may be inaccurate if there are concurrent kernels running. (?)

I do think that the overhead of NVTX is nearly zero when nsys or other tools are not attached. We use it and I haven't noticed a penalty. Typically we wrap up the calls in utility functions that we have the option of disabling with a preprocessor definition.

[gaoteng-git]

You may find these CUDA event benchmarks useful: https://github.com/harrism/cuda_event_benchmark. In my tests on V100 / Ubuntu 18.04 / CUDA 10.2 / AMD Ryzen 7 3700, a default event record (timing enable) has a throughput of about 400K records per second. With timing disabled it is 10x faster, but if you are intending to use events for timing, go with the lower throughput.

The event approach works well with tools like Google Benchmark, however you may need to take extra steps to flush the L2 cache between kernels if you need to benchmark the performance assuming a cold cache. You can see how we do this in RAPIDS libcudf benchmarks with this class, and an example benchmark that uses it:
https://github.com/rapidsai/cudf/blob/f78f80e94c74c08fface696cfd7e03881b9b0380/cpp/benchmarks/transpose/transpose_benchmark.cu#L46-L49

Note that using CUDA events for timing may be inaccurate if there are concurrent kernels running. (?)

I do think that the overhead of NVTX is nearly zero when nsys or other tools are not attached. We use it and I haven't noticed a penalty. Typically we wrap up the calls in utility functions that we have the option of disabling with a preprocessor definition.

I'm shock by your show up! Your developer blogs in early years are really helpful and I learnt a lot from them. Thank you very much!
When concurrent kernels are running, the elapsed time that cuda events record on the same kernel code may be very different in different running tries, because of other concurrent kernels' competition on the same GPU cores resource. Is my understanding right?
Thanks a lot again for your help!

[harrism]

No worries, happy to help. I think your understanding is correct.