Hybrid MPI and CUDA Parallelization for CFD Applications on Multi-GPU HPC Clusters

Algorithm 1

Parallel algorithm for CFD on multi-GPU HPC clusters.

(1)	MPI_Init (&argc, &argv);
(2)	Device_Query ( );
(3)	cudaMemcpy (d_a, h_a, sizeof(float)n, cudaMemcpyHostToDevice);
(4)	//Kernel execution start
(5)	for i = 0; i < max_step; i++
(6)	Boundary_Processing_GPU<<<Block_size, Thread_size>>> ( );
(7)	Time_Step_GPU<<<Block_size, Thread_size>>> ();
(8)	Primitive_Variables_Exchange ();
(9)	Grad_Primitive_Variables_GPU<<<Block_size, Thread_size>>> ( );
(10)	Grad_primitive_Variables_Exchange ( );
(11)	Flux_GPU<<<Block_size, Thread_size>>> ( );
(12)	Primitive_Variables_Update_GPU<<<Block_size, Thread_size>>> ( );
(13)	end for
(14)	//kernel execution end
(15)	cudaMemcpy (h_a, d_a, sizeof(float)n. cudaMemcpyDeviceToHost);
(16)	Flow_post-processing ( );
(17)	MPI_finalize ( );