Allocate memory for two cards - a large run-time code

Accelerated Computing CUDA CUDA Programming and Performance
1

I have the code:

time_start_block = clock();

	cudaSetDevice(0);

	float *matrix_device_0;

	cudaMalloc((void **) &matrix_device_0, N * N * sizeof(float) );

	float *phi_device_0;

	cudaMalloc((void **) &phi_device_0, N * sizeof(float) );

	float *vector_TEMP_device_0;

	cudaMalloc((void **) &vector_TEMP_device_0, N * sizeof(float) );

	cudaMemcpy(matrix_device_0, matrix_A, N * N * sizeof(float), cudaMemcpyHostToDevice);

	

	cudaSetDevice(1);

	float *matrix_device_1;

	cudaMalloc((void **) &matrix_device_1, N * N * sizeof(float) );

	float *phi_device_1;

	cudaMalloc((void **) &phi_device_1, N * sizeof(float) );

	float *vector_TEMP_device_1;

	cudaMalloc((void **) &vector_TEMP_device_1, N * sizeof(float) );

	

	cudaMemcpy(matrix_device_1, matrix_B, N * N * sizeof(float), cudaMemcpyHostToDevice);

	time_end_block = clock();

	time_dif_block = (float)(time_end_block-time_start_block) / CLOCKS_PER_SEC;

The measured execution time is approximately 45 seconds (Tesla S1070).

Why it takes so long?

2

I solved the problem.

Physically, the computer has only 4 GB of RAM. This program will allocate a lot of memory and use the SWAP.

When I called ‘cudaMemcpy’ program copied the data from the SWAP memory. It really slowed down.