1 2 static char help[] = "Test PC redistribute on matrix with load imbalance. \n\ 3 Modified from src/ksp/ksp/tutorials/ex2.c.\n\ 4 Input parameters include:\n\ 5 -random_exact_sol : use a random exact solution vector\n\ 6 -view_exact_sol : write exact solution vector to stdout\n\ 7 -n <mesh_y> : number of mesh points\n\n"; 8 /* 9 Example: 10 mpiexec -n 8 ./ex3 -n 10000 -ksp_type cg -pc_type bjacobi -sub_pc_type icc -ksp_rtol 1.e-8 -log_view 11 mpiexec -n 8 ./ex3 -n 10000 -ksp_type preonly -pc_type redistribute -redistribute_ksp_type cg -redistribute_pc_type bjacobi -redistribute_sub_pc_type icc -redistribute_ksp_rtol 1.e-8 -log_view 12 */ 13 14 #include <petscksp.h> 15 16 int main(int argc,char **args) 17 { 18 Vec x,b,u; /* approx solution, RHS, exact solution */ 19 Mat A; /* linear system matrix */ 20 KSP ksp; /* linear solver context */ 21 PetscRandom rctx; /* random number generator context */ 22 PetscReal norm; /* norm of solution error */ 23 PetscInt i,j,Ii,J,Istart,Iend,m,n = 7,its,nloc,matdistribute=0; 24 PetscBool flg = PETSC_FALSE; 25 PetscScalar v; 26 PetscMPIInt rank,size; 27 #if defined(PETSC_USE_LOG) 28 PetscLogStage stage; 29 #endif 30 31 PetscFunctionBeginUser; 32 PetscCall(PetscInitialize(&argc,&args,(char*)0,help)); 33 PetscCallMPI(MPI_Comm_size(PETSC_COMM_WORLD,&size)); 34 PetscCallMPI(MPI_Comm_rank(PETSC_COMM_WORLD,&rank)); 35 PetscCheck(size > 1,PETSC_COMM_WORLD,PETSC_ERR_WRONG_MPI_SIZE,"This example requires at least 2 MPI processes!"); 36 37 PetscCall(PetscOptionsGetInt(NULL,NULL,"-n",&n,NULL)); 38 PetscCall(PetscOptionsGetInt(NULL,NULL,"-matdistribute",&matdistribute,NULL)); 39 /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 40 Compute the matrix and right-hand-side vector that define 41 the linear system, Ax = b. 42 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ 43 switch(matdistribute) { 44 case 1: /* very imbalanced process load for matrix A */ 45 m = (1+size)*size; 46 nloc = (rank+1)*n; 47 if (rank == size-1) { /* proc[size-1] stores all remaining rows */ 48 nloc = m*n; 49 for (i=0; i<size-1; i++) { 50 nloc -= (i+1)*n; 51 } 52 } 53 break; 54 default: /* proc[0] and proc[1] load much smaller row blocks, the rest processes have same loads */ 55 if (rank == 0 || rank == 1) { 56 nloc = n; 57 } else { 58 nloc = 10*n; /* 10x larger load */ 59 } 60 m = 2 + (size-2)*10; 61 break; 62 } 63 PetscCall(MatCreate(PETSC_COMM_WORLD,&A)); 64 PetscCall(MatSetSizes(A,nloc,nloc,PETSC_DECIDE,PETSC_DECIDE)); 65 PetscCall(MatSetFromOptions(A)); 66 PetscCall(MatMPIAIJSetPreallocation(A,5,NULL,5,NULL)); 67 PetscCall(MatSeqAIJSetPreallocation(A,5,NULL)); 68 PetscCall(MatSetUp(A)); 69 70 PetscCall(MatGetOwnershipRange(A,&Istart,&Iend)); 71 nloc = Iend-Istart; 72 PetscCall(PetscSynchronizedPrintf(PETSC_COMM_WORLD,"[%d] A Istart,Iend: %" PetscInt_FMT " %" PetscInt_FMT "; nloc %" PetscInt_FMT "\n",rank,Istart,Iend,nloc)); 73 PetscCall(PetscSynchronizedFlush(PETSC_COMM_WORLD,PETSC_STDOUT)); 74 75 PetscCall(PetscLogStageRegister("Assembly", &stage)); 76 PetscCall(PetscLogStagePush(stage)); 77 for (Ii=Istart; Ii<Iend; Ii++) { 78 v = -1.0; i = Ii/n; j = Ii - i*n; 79 if (i>0) {J = Ii - n; PetscCall(MatSetValues(A,1,&Ii,1,&J,&v,INSERT_VALUES));} 80 if (i<m-1) {J = Ii + n; PetscCall(MatSetValues(A,1,&Ii,1,&J,&v,INSERT_VALUES));} 81 if (j>0) {J = Ii - 1; PetscCall(MatSetValues(A,1,&Ii,1,&J,&v,INSERT_VALUES));} 82 if (j<n-1) {J = Ii + 1; PetscCall(MatSetValues(A,1,&Ii,1,&J,&v,INSERT_VALUES));} 83 v = 4.0; PetscCall(MatSetValues(A,1,&Ii,1,&Ii,&v,INSERT_VALUES)); 84 } 85 PetscCall(MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY)); 86 PetscCall(MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY)); 87 PetscCall(PetscLogStagePop()); 88 89 /* A is symmetric. Set symmetric flag to enable ICC/Cholesky preconditioner */ 90 PetscCall(MatSetOption(A,MAT_SYMMETRIC,PETSC_TRUE)); 91 92 /* Create parallel vectors. */ 93 PetscCall(VecCreate(PETSC_COMM_WORLD,&u)); 94 PetscCall(VecSetSizes(u,nloc,PETSC_DECIDE)); 95 PetscCall(VecSetFromOptions(u)); 96 PetscCall(VecDuplicate(u,&b)); 97 PetscCall(VecDuplicate(b,&x)); 98 99 /* Set exact solution; then compute right-hand-side vector. */ 100 PetscCall(PetscOptionsGetBool(NULL,NULL,"-random_exact_sol",&flg,NULL)); 101 if (flg) { 102 PetscCall(PetscRandomCreate(PETSC_COMM_WORLD,&rctx)); 103 PetscCall(PetscRandomSetFromOptions(rctx)); 104 PetscCall(VecSetRandom(u,rctx)); 105 PetscCall(PetscRandomDestroy(&rctx)); 106 } else { 107 PetscCall(VecSet(u,1.0)); 108 } 109 PetscCall(MatMult(A,u,b)); 110 111 /* View the exact solution vector if desired */ 112 flg = PETSC_FALSE; 113 PetscCall(PetscOptionsGetBool(NULL,NULL,"-view_exact_sol",&flg,NULL)); 114 if (flg) PetscCall(VecView(u,PETSC_VIEWER_STDOUT_WORLD)); 115 116 /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 117 Create the linear solver and set various options 118 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ 119 PetscCall(KSPCreate(PETSC_COMM_WORLD,&ksp)); 120 PetscCall(KSPSetOperators(ksp,A,A)); 121 PetscCall(KSPSetTolerances(ksp,1.e-2/((m+1)*(n+1)),PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT)); 122 PetscCall(KSPSetFromOptions(ksp)); 123 124 /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 125 Solve the linear system 126 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ 127 PetscCall(KSPSolve(ksp,b,x)); 128 129 /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 130 Check solution and clean up 131 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ 132 PetscCall(VecAXPY(x,-1.0,u)); 133 PetscCall(VecNorm(x,NORM_2,&norm)); 134 PetscCall(KSPGetIterationNumber(ksp,&its)); 135 PetscCall(PetscPrintf(PETSC_COMM_WORLD,"Norm of error %g iterations %" PetscInt_FMT "\n",(double)norm,its)); 136 137 /* Free work space. */ 138 PetscCall(KSPDestroy(&ksp)); 139 PetscCall(VecDestroy(&u)); PetscCall(VecDestroy(&x)); 140 PetscCall(VecDestroy(&b)); PetscCall(MatDestroy(&A)); 141 PetscCall(PetscFinalize()); 142 return 0; 143 } 144 145 /*TEST 146 147 test: 148 nsize: 8 149 args: -n 100 -ksp_type cg -pc_type bjacobi -sub_pc_type icc -ksp_rtol 1.e-8 150 151 test: 152 suffix: 2 153 nsize: 8 154 args: -n 100 -ksp_type preonly -pc_type redistribute -redistribute_ksp_type cg -redistribute_pc_type bjacobi -redistribute_sub_pc_type icc -redistribute_ksp_rtol 1.e-8 155 156 TEST*/ 157