Issue on parallelising memcpy

[BrownZer]

Dear experts,

I’m a beginner in CUDA programming.
I was following the to perform a parallelised host-to-device memcpy and I wrote the example code below.
However when I run the nsight systems “trace” I found that the memcpy’s don’t seem to be parallelised, as shown in the attached picture.

May i ask what was missed in the example code or this can be done by other approaches?

Thanks ahead!
Brown

	void Routine()
	{
		cudaStream_t s1;
		cudaStream_t s2;

		CUDA_CHECK(cudaStreamCreate(&s1));
		CUDA_CHECK(cudaStreamCreate(&s2));

		double* host_1;
		double* host_2;

		CUDA_CHECK(cudaHostAlloc(&host_1, MYSIZE, cudaHostAllocDefault));
		CUDA_CHECK(cudaHostAlloc(&host_2, MYSIZE, cudaHostAllocDefault));

		double* device_1;
		double* device_2;

		CUDA_CHECK(cudaMallocAsync(&device_1, MYSIZE, s1));
		CUDA_CHECK(cudaMallocAsync(&device_2, MYSIZE, s2));

		CUDA_CHECK(cudaMemcpyAsync(device_1, host_1, MYSIZE, cudaMemcpyHostToDevice, s1));
		CUDA_CHECK(cudaMemcpyAsync(device_2, host_2, MYSIZE, cudaMemcpyHostToDevice, s2));

		// clean up
		CUDA_CHECK(cudaFreeHost(host_1));
		CUDA_CHECK(cudaFreeHost(host_2));
		CUDA_CHECK(cudaFreeAsync(device_1, s1));
		CUDA_CHECK(cudaFreeAsync(device_2, s2));
		CUDA_CHECK(cudaStreamDestroy(s1));
		CUDA_CHECK(cudaStreamDestroy(s2));
	}

[Robert_Crovella]

There isn’t any expectation that such operations would execute “in parallel”. The pipe that connects the host to the device has a particular finite bandwidth.

No, it can’t be done by other approaches. That is, there is nothing you can do to force the two operations to begin at the same time and run concurrently.

[BrownZer]

Thanks a lot for the reply! It clearly answers my question :)