Actual source code: pcmpi.c
1: /*
2: This file creates an MPI parallel KSP from a sequential PC that lives on MPI rank 0.
3: It is intended to allow using PETSc MPI parallel linear solvers from non-MPI codes.
5: That program may use OpenMP to compute the right-hand side and matrix for the linear system
7: The code uses MPI_COMM_WORLD below but maybe it should be PETSC_COMM_WORLD
9: The resulting KSP and PC can only be controlled via the options database, though some common commands
10: could be passed through the server.
12: */
13: #include <petsc/private/pcimpl.h>
14: #include <petsc/private/kspimpl.h>
15: #include <petscts.h>
16: #include <petsctao.h>
17: #if defined(PETSC_HAVE_PTHREAD_MUTEX)
18: #include <pthread.h>
19: #endif
21: #define PC_MPI_MAX_RANKS 256
22: #define PC_MPI_COMM_WORLD MPI_COMM_WORLD
24: typedef struct {
25: KSP ksps[PC_MPI_MAX_RANKS]; /* The addresses of the MPI parallel KSP on each process, NULL when not on a process. */
26: PetscMPIInt sendcount[PC_MPI_MAX_RANKS], displ[PC_MPI_MAX_RANKS]; /* For scatter/gather of rhs/solution */
27: PetscMPIInt NZ[PC_MPI_MAX_RANKS], NZdispl[PC_MPI_MAX_RANKS]; /* For scatter of nonzero values in matrix (and nonzero column indices initially */
28: PetscInt mincntperrank; /* minimum number of desired matrix rows per active rank in MPI parallel KSP solve */
29: PetscBool alwaysuseserver; /* for debugging use the server infrastructure even if only one MPI process is used for the solve */
30: } PC_MPI;
32: typedef enum {
33: PCMPI_EXIT, /* exit the PC server loop, means the controlling sequential program is done */
34: PCMPI_CREATE,
35: PCMPI_SET_MAT, /* set original matrix (or one with different nonzero pattern) */
36: PCMPI_UPDATE_MAT_VALUES, /* update current matrix with new nonzero values */
37: PCMPI_SOLVE,
38: PCMPI_VIEW,
39: PCMPI_DESTROY /* destroy a PC that is no longer needed */
40: } PCMPICommand;
42: static MPI_Comm PCMPIComms[PC_MPI_MAX_RANKS];
43: static PetscBool PCMPICommSet = PETSC_FALSE;
44: static PetscInt PCMPISolveCounts[PC_MPI_MAX_RANKS], PCMPIKSPCounts[PC_MPI_MAX_RANKS], PCMPIMatCounts[PC_MPI_MAX_RANKS], PCMPISolveCountsSeq = 0, PCMPIKSPCountsSeq = 0;
45: static PetscInt PCMPIIterations[PC_MPI_MAX_RANKS], PCMPISizes[PC_MPI_MAX_RANKS], PCMPIIterationsSeq = 0, PCMPISizesSeq = 0;
46: static PetscLogEvent EventServerDist, EventServerDistMPI;
47: #if defined(PETSC_HAVE_PTHREAD_MUTEX)
48: static pthread_mutex_t *PCMPIServerLocks;
49: #else
50: static void *PCMPIServerLocks;
51: #endif
53: static PetscErrorCode PCMPICommsCreate(void)
54: {
55: MPI_Comm comm = PC_MPI_COMM_WORLD;
56: PetscMPIInt size, rank, i;
58: PetscFunctionBegin;
59: PetscCallMPI(MPI_Comm_size(comm, &size));
60: PetscCheck(size <= PC_MPI_MAX_RANKS, PETSC_COMM_SELF, PETSC_ERR_SUP, "No support for using more than PC_MPI_MAX_RANKS MPI ranks in an MPI linear solver server solve");
61: PetscCallMPI(MPI_Comm_rank(comm, &rank));
62: /* comm for size 1 is useful only for debugging */
63: for (i = 0; i < size; i++) {
64: PetscMPIInt color = rank < i + 1 ? 0 : MPI_UNDEFINED;
65: PetscCallMPI(MPI_Comm_split(comm, color, 0, &PCMPIComms[i]));
66: PCMPISolveCounts[i] = 0;
67: PCMPIKSPCounts[i] = 0;
68: PCMPIIterations[i] = 0;
69: PCMPISizes[i] = 0;
70: }
71: PCMPICommSet = PETSC_TRUE;
72: PetscFunctionReturn(PETSC_SUCCESS);
73: }
75: static PetscErrorCode PCMPICommsDestroy(void)
76: {
77: MPI_Comm comm = PC_MPI_COMM_WORLD;
78: PetscMPIInt size, rank, i;
80: PetscFunctionBegin;
81: if (!PCMPICommSet) PetscFunctionReturn(PETSC_SUCCESS);
82: PetscCallMPI(MPI_Comm_size(comm, &size));
83: PetscCallMPI(MPI_Comm_rank(comm, &rank));
84: for (i = 0; i < size; i++) {
85: if (PCMPIComms[i] != MPI_COMM_NULL) PetscCallMPI(MPI_Comm_free(&PCMPIComms[i]));
86: }
87: PCMPICommSet = PETSC_FALSE;
88: PetscFunctionReturn(PETSC_SUCCESS);
89: }
91: static PetscErrorCode PCMPICreate(PC pc)
92: {
93: PC_MPI *km = pc ? (PC_MPI *)pc->data : NULL;
94: MPI_Comm comm = PC_MPI_COMM_WORLD;
95: KSP ksp;
96: PetscInt N[2], mincntperrank = 0;
97: PetscMPIInt size;
98: Mat sA;
99: char *cprefix = NULL;
100: PetscMPIInt len = 0;
102: PetscFunctionBegin;
103: PCMPIServerInSolve = PETSC_TRUE;
104: if (!PCMPICommSet) PetscCall(PCMPICommsCreate());
105: PetscCallMPI(MPI_Comm_size(comm, &size));
106: if (pc) {
107: if (size == 1) PetscCall(PetscPrintf(PETSC_COMM_SELF, "Warning: Running KSP type of MPI on a one rank MPI run, this will be less efficient then not using this type\n"));
108: PetscCall(PCGetOperators(pc, &sA, &sA));
109: PetscCall(MatGetSize(sA, &N[0], &N[1]));
110: }
111: PetscCallMPI(MPI_Bcast(N, 2, MPIU_INT, 0, comm));
113: /* choose a suitable sized MPI_Comm for the problem to be solved on */
114: if (km) mincntperrank = km->mincntperrank;
115: PetscCallMPI(MPI_Bcast(&mincntperrank, 1, MPI_INT, 0, comm));
116: comm = PCMPIComms[PetscMin(size, PetscMax(1, N[0] / mincntperrank)) - 1];
117: if (comm == MPI_COMM_NULL) {
118: ksp = NULL;
119: PCMPIServerInSolve = PETSC_FALSE;
120: PetscFunctionReturn(PETSC_SUCCESS);
121: }
122: PetscCall(PetscLogStagePush(PCMPIStage));
123: PetscCall(KSPCreate(comm, &ksp));
124: PetscCall(KSPSetNestLevel(ksp, 1));
125: PetscCall(PetscObjectSetTabLevel((PetscObject)ksp, 1));
126: PetscCall(PetscLogStagePop());
127: PetscCallMPI(MPI_Gather(&ksp, 1, MPI_AINT, pc ? km->ksps : NULL, 1, MPI_AINT, 0, comm));
128: if (pc) {
129: size_t slen;
130: const char *prefix = NULL;
131: char *found = NULL;
133: PetscCallMPI(MPI_Comm_size(comm, &size));
134: PCMPIKSPCounts[size - 1]++;
135: /* Created KSP gets prefix of PC minus the mpi_linear_solver_server_ portion */
136: PetscCall(PCGetOptionsPrefix(pc, &prefix));
137: PetscCheck(prefix, PETSC_COMM_SELF, PETSC_ERR_PLIB, "PCMPI missing required prefix");
138: PetscCall(PetscStrallocpy(prefix, &cprefix));
139: PetscCall(PetscStrstr(cprefix, "mpi_linear_solver_server_", &found));
140: PetscCheck(found, PETSC_COMM_SELF, PETSC_ERR_PLIB, "PCMPI missing mpi_linear_solver_server_ portion of prefix");
141: *found = 0;
142: PetscCall(PetscStrlen(cprefix, &slen));
143: len = (PetscMPIInt)slen;
144: }
145: PetscCallMPI(MPI_Bcast(&len, 1, MPI_INT, 0, comm));
146: if (len) {
147: if (!pc) PetscCall(PetscMalloc1(len + 1, &cprefix));
148: PetscCallMPI(MPI_Bcast(cprefix, len + 1, MPI_CHAR, 0, comm));
149: PetscCall(KSPSetOptionsPrefix(ksp, cprefix));
150: }
151: PetscCall(PetscFree(cprefix));
152: PCMPIServerInSolve = PETSC_FALSE;
153: PetscFunctionReturn(PETSC_SUCCESS);
154: }
156: static PetscErrorCode PCMPISetMat(PC pc)
157: {
158: PC_MPI *km = pc ? (PC_MPI *)pc->data : NULL;
159: Mat A;
160: PetscInt m, n, j, bs;
161: Mat sA;
162: MPI_Comm comm = PC_MPI_COMM_WORLD;
163: KSP ksp;
164: PetscLayout layout;
165: const PetscInt *IA = NULL, *JA = NULL, *ia, *ja;
166: const PetscInt *range;
167: PetscMPIInt *NZ = NULL, sendcounti[PC_MPI_MAX_RANKS], displi[PC_MPI_MAX_RANKS], *NZdispl = NULL, nz, size, i;
168: const PetscScalar *a = NULL, *sa;
169: PetscInt matproperties[8] = {0}, rstart, rend;
170: char *cprefix;
172: PetscFunctionBegin;
173: PetscCallMPI(MPI_Scatter(pc ? km->ksps : NULL, 1, MPI_AINT, &ksp, 1, MPI_AINT, 0, comm));
174: if (!ksp) PetscFunctionReturn(PETSC_SUCCESS);
175: PCMPIServerInSolve = PETSC_TRUE;
176: PetscCall(PetscLogEventBegin(EventServerDist, NULL, NULL, NULL, NULL));
177: PetscCall(PetscObjectGetComm((PetscObject)ksp, &comm));
178: if (pc) {
179: PetscBool isset, issymmetric, ishermitian, isspd, isstructurallysymmetric;
180: const char *prefix;
181: size_t clen;
183: PetscCallMPI(MPI_Comm_size(comm, &size));
184: PCMPIMatCounts[size - 1]++;
185: PetscCall(PCGetOperators(pc, &sA, &sA));
186: PetscCall(MatGetSize(sA, &matproperties[0], &matproperties[1]));
187: PetscCall(MatGetBlockSize(sA, &bs));
188: matproperties[2] = bs;
189: PetscCall(MatIsSymmetricKnown(sA, &isset, &issymmetric));
190: matproperties[3] = !isset ? 0 : (issymmetric ? 1 : 2);
191: PetscCall(MatIsHermitianKnown(sA, &isset, &ishermitian));
192: matproperties[4] = !isset ? 0 : (ishermitian ? 1 : 2);
193: PetscCall(MatIsSPDKnown(sA, &isset, &isspd));
194: matproperties[5] = !isset ? 0 : (isspd ? 1 : 2);
195: PetscCall(MatIsStructurallySymmetricKnown(sA, &isset, &isstructurallysymmetric));
196: matproperties[6] = !isset ? 0 : (isstructurallysymmetric ? 1 : 2);
197: /* Created Mat gets prefix of input Mat PLUS the mpi_linear_solver_server_ portion */
198: PetscCall(MatGetOptionsPrefix(sA, &prefix));
199: PetscCall(PetscStrallocpy(prefix, &cprefix));
200: PetscCall(PetscStrlen(cprefix, &clen));
201: matproperties[7] = (PetscInt)clen;
202: }
203: PetscCallMPI(MPI_Bcast(matproperties, PETSC_STATIC_ARRAY_LENGTH(matproperties), MPIU_INT, 0, comm));
205: /* determine ownership ranges of matrix columns */
206: PetscCall(PetscLayoutCreate(comm, &layout));
207: PetscCall(PetscLayoutSetBlockSize(layout, matproperties[2]));
208: PetscCall(PetscLayoutSetSize(layout, matproperties[1]));
209: PetscCall(PetscLayoutSetUp(layout));
210: PetscCall(PetscLayoutGetLocalSize(layout, &n));
211: PetscCall(PetscLayoutDestroy(&layout));
213: /* determine ownership ranges of matrix rows */
214: PetscCall(PetscLayoutCreate(comm, &layout));
215: PetscCall(PetscLayoutSetBlockSize(layout, matproperties[2]));
216: PetscCall(PetscLayoutSetSize(layout, matproperties[0]));
217: PetscCall(PetscLayoutSetUp(layout));
218: PetscCall(PetscLayoutGetLocalSize(layout, &m));
219: PetscCall(PetscLayoutGetRange(layout, &rstart, &rend));
221: PetscCall(PetscLogEventBegin(EventServerDistMPI, NULL, NULL, NULL, NULL));
222: /* copy over the matrix nonzero structure and values */
223: if (pc) {
224: PetscCall(MatGetRowIJ(sA, 0, PETSC_FALSE, PETSC_FALSE, NULL, &IA, &JA, NULL));
225: if (!PCMPIServerUseShmget) {
226: NZ = km->NZ;
227: NZdispl = km->NZdispl;
228: PetscCall(PetscLayoutGetRanges(layout, &range));
229: for (i = 0; i < size; i++) {
230: sendcounti[i] = (PetscMPIInt)(1 + range[i + 1] - range[i]);
231: NZ[i] = (PetscMPIInt)(IA[range[i + 1]] - IA[range[i]]);
232: }
233: displi[0] = 0;
234: NZdispl[0] = 0;
235: for (j = 1; j < size; j++) {
236: displi[j] = displi[j - 1] + sendcounti[j - 1] - 1;
237: NZdispl[j] = NZdispl[j - 1] + NZ[j - 1];
238: }
239: }
240: PetscCall(MatSeqAIJGetArrayRead(sA, &sa));
241: }
242: PetscCall(PetscLayoutDestroy(&layout));
244: PetscCall(MatCreate(comm, &A));
245: if (matproperties[7] > 0) {
246: if (!pc) PetscCall(PetscMalloc1(matproperties[7] + 1, &cprefix));
247: PetscCallMPI(MPI_Bcast(cprefix, matproperties[7] + 1, MPI_CHAR, 0, comm));
248: PetscCall(MatSetOptionsPrefix(A, cprefix));
249: PetscCall(PetscFree(cprefix));
250: }
251: PetscCall(MatAppendOptionsPrefix(A, "mpi_linear_solver_server_"));
252: PetscCall(MatSetSizes(A, m, n, matproperties[0], matproperties[1]));
253: PetscCall(MatSetType(A, MATMPIAIJ));
255: if (!PCMPIServerUseShmget) {
256: PetscCallMPI(MPI_Scatter(NZ, 1, MPI_INT, &nz, 1, MPI_INT, 0, comm));
257: PetscCall(PetscMalloc3(n + 1, &ia, nz, &ja, nz, &a));
258: PetscCallMPI(MPI_Scatterv(IA, sendcounti, displi, MPIU_INT, (void *)ia, n + 1, MPIU_INT, 0, comm));
259: PetscCallMPI(MPI_Scatterv(JA, NZ, NZdispl, MPIU_INT, (void *)ja, nz, MPIU_INT, 0, comm));
260: PetscCallMPI(MPI_Scatterv(sa, NZ, NZdispl, MPIU_SCALAR, (void *)a, nz, MPIU_SCALAR, 0, comm));
261: } else {
262: const void *addr[3] = {(const void **)IA, (const void **)JA, (const void **)sa};
263: PCMPIServerAddresses *addresses;
265: PetscCall(PetscNew(&addresses));
266: addresses->n = 3;
267: PetscCall(PetscShmgetMapAddresses(comm, addresses->n, addr, addresses->addr));
268: ia = rstart + (PetscInt *)addresses->addr[0];
269: ja = ia[0] + (PetscInt *)addresses->addr[1];
270: a = ia[0] + (PetscScalar *)addresses->addr[2];
271: PetscCall(PetscObjectContainerCompose((PetscObject)A, "PCMPIServerAddresses", (void *)addresses, (PetscErrorCode(*)(void *))PCMPIServerAddressesDestroy));
272: }
274: if (pc) {
275: PetscCall(MatSeqAIJRestoreArrayRead(sA, &sa));
276: PetscCall(MatRestoreRowIJ(sA, 0, PETSC_FALSE, PETSC_FALSE, NULL, &IA, &JA, NULL));
277: }
278: PetscCall(PetscLogEventEnd(EventServerDistMPI, NULL, NULL, NULL, NULL));
280: PetscCall(PetscLogStagePush(PCMPIStage));
281: PetscCall(MatMPIAIJSetPreallocationCSR(A, ia, ja, a));
282: PetscCall(MatSetBlockSize(A, matproperties[2]));
284: if (matproperties[3]) PetscCall(MatSetOption(A, MAT_SYMMETRIC, matproperties[3] == 1 ? PETSC_TRUE : PETSC_FALSE));
285: if (matproperties[4]) PetscCall(MatSetOption(A, MAT_HERMITIAN, matproperties[4] == 1 ? PETSC_TRUE : PETSC_FALSE));
286: if (matproperties[5]) PetscCall(MatSetOption(A, MAT_SPD, matproperties[5] == 1 ? PETSC_TRUE : PETSC_FALSE));
287: if (matproperties[6]) PetscCall(MatSetOption(A, MAT_STRUCTURALLY_SYMMETRIC, matproperties[6] == 1 ? PETSC_TRUE : PETSC_FALSE));
289: if (!PCMPIServerUseShmget) PetscCall(PetscFree3(ia, ja, a));
290: PetscCall(KSPSetOperators(ksp, A, A));
291: if (!ksp->vec_sol) PetscCall(MatCreateVecs(A, &ksp->vec_sol, &ksp->vec_rhs));
292: PetscCall(PetscLogStagePop());
293: if (pc && !PCMPIServerUseShmget) { /* needed for scatterv/gatherv of rhs and solution */
294: const PetscInt *range;
296: PetscCall(VecGetOwnershipRanges(ksp->vec_sol, &range));
297: for (i = 0; i < size; i++) {
298: km->sendcount[i] = (PetscMPIInt)(range[i + 1] - range[i]);
299: km->displ[i] = (PetscMPIInt)range[i];
300: }
301: }
302: PetscCall(MatDestroy(&A));
303: PetscCall(PetscLogEventEnd(EventServerDist, NULL, NULL, NULL, NULL));
304: PetscCall(KSPSetFromOptions(ksp));
305: PCMPIServerInSolve = PETSC_FALSE;
306: PetscFunctionReturn(PETSC_SUCCESS);
307: }
309: static PetscErrorCode PCMPIUpdateMatValues(PC pc)
310: {
311: PC_MPI *km = pc ? (PC_MPI *)pc->data : NULL;
312: KSP ksp;
313: Mat sA, A;
314: MPI_Comm comm = PC_MPI_COMM_WORLD;
315: const PetscInt *ia, *IA;
316: const PetscScalar *a;
317: PetscCount nz;
318: const PetscScalar *sa = NULL;
319: PetscMPIInt size;
320: PetscInt rstart, matproperties[4] = {0, 0, 0, 0};
322: PetscFunctionBegin;
323: if (pc) {
324: PetscCall(PCGetOperators(pc, &sA, &sA));
325: PetscCall(MatSeqAIJGetArrayRead(sA, &sa));
326: PetscCall(MatGetRowIJ(sA, 0, PETSC_FALSE, PETSC_FALSE, NULL, &IA, NULL, NULL));
327: }
328: PetscCallMPI(MPI_Scatter(pc ? km->ksps : NULL, 1, MPI_AINT, &ksp, 1, MPI_AINT, 0, comm));
329: if (!ksp) PetscFunctionReturn(PETSC_SUCCESS);
330: PCMPIServerInSolve = PETSC_TRUE;
331: PetscCall(PetscLogEventBegin(EventServerDist, NULL, NULL, NULL, NULL));
332: PetscCall(PetscObjectGetComm((PetscObject)ksp, &comm));
333: PetscCallMPI(MPI_Comm_size(comm, &size));
334: PCMPIMatCounts[size - 1]++;
335: PetscCall(KSPGetOperators(ksp, NULL, &A));
336: PetscCall(PetscLogEventBegin(EventServerDistMPI, NULL, NULL, NULL, NULL));
337: if (!PCMPIServerUseShmget) {
338: PetscCall(MatMPIAIJGetNumberNonzeros(A, &nz));
339: PetscCall(PetscMalloc1(nz, &a));
340: PetscCallMPI(MPI_Scatterv(sa, pc ? km->NZ : NULL, pc ? km->NZdispl : NULL, MPIU_SCALAR, (void *)a, nz, MPIU_SCALAR, 0, comm));
341: } else {
342: PetscCall(MatGetOwnershipRange(A, &rstart, NULL));
343: PCMPIServerAddresses *addresses;
344: PetscCall(PetscObjectContainerQuery((PetscObject)A, "PCMPIServerAddresses", (void **)&addresses));
345: ia = rstart + (PetscInt *)addresses->addr[0];
346: a = ia[0] + (PetscScalar *)addresses->addr[2];
347: }
348: PetscCall(PetscLogEventEnd(EventServerDistMPI, NULL, NULL, NULL, NULL));
349: if (pc) {
350: PetscBool isset, issymmetric, ishermitian, isspd, isstructurallysymmetric;
352: PetscCall(MatSeqAIJRestoreArrayRead(sA, &sa));
353: PetscCall(MatRestoreRowIJ(sA, 0, PETSC_FALSE, PETSC_FALSE, NULL, &IA, NULL, NULL));
355: PetscCall(MatIsSymmetricKnown(sA, &isset, &issymmetric));
356: matproperties[0] = !isset ? 0 : (issymmetric ? 1 : 2);
357: PetscCall(MatIsHermitianKnown(sA, &isset, &ishermitian));
358: matproperties[1] = !isset ? 0 : (ishermitian ? 1 : 2);
359: PetscCall(MatIsSPDKnown(sA, &isset, &isspd));
360: matproperties[2] = !isset ? 0 : (isspd ? 1 : 2);
361: PetscCall(MatIsStructurallySymmetricKnown(sA, &isset, &isstructurallysymmetric));
362: matproperties[3] = !isset ? 0 : (isstructurallysymmetric ? 1 : 2);
363: }
364: PetscCall(MatUpdateMPIAIJWithArray(A, a));
365: if (!PCMPIServerUseShmget) PetscCall(PetscFree(a));
366: PetscCallMPI(MPI_Bcast(matproperties, 4, MPIU_INT, 0, comm));
367: /* if any of these properties was previously set and is now not set this will result in incorrect properties in A since there is no way to unset a property */
368: if (matproperties[0]) PetscCall(MatSetOption(A, MAT_SYMMETRIC, matproperties[0] == 1 ? PETSC_TRUE : PETSC_FALSE));
369: if (matproperties[1]) PetscCall(MatSetOption(A, MAT_HERMITIAN, matproperties[1] == 1 ? PETSC_TRUE : PETSC_FALSE));
370: if (matproperties[2]) PetscCall(MatSetOption(A, MAT_SPD, matproperties[2] == 1 ? PETSC_TRUE : PETSC_FALSE));
371: if (matproperties[3]) PetscCall(MatSetOption(A, MAT_STRUCTURALLY_SYMMETRIC, matproperties[3] == 1 ? PETSC_TRUE : PETSC_FALSE));
372: PetscCall(PetscLogEventEnd(EventServerDist, NULL, NULL, NULL, NULL));
373: PCMPIServerInSolve = PETSC_FALSE;
374: PetscFunctionReturn(PETSC_SUCCESS);
375: }
377: static PetscErrorCode PCMPISolve(PC pc, Vec B, Vec X)
378: {
379: PC_MPI *km = pc ? (PC_MPI *)pc->data : NULL;
380: KSP ksp;
381: MPI_Comm comm = PC_MPI_COMM_WORLD;
382: const PetscScalar *sb = NULL, *x;
383: PetscScalar *b, *sx = NULL;
384: PetscInt its, n;
385: PetscMPIInt size;
386: void *addr[2];
388: PetscFunctionBegin;
389: PetscCallMPI(MPI_Scatter(pc ? km->ksps : &ksp, 1, MPI_AINT, &ksp, 1, MPI_AINT, 0, comm));
390: if (!ksp) PetscFunctionReturn(PETSC_SUCCESS);
391: PCMPIServerInSolve = PETSC_TRUE;
392: PetscCall(PetscLogEventBegin(EventServerDist, NULL, NULL, NULL, NULL));
393: PetscCall(PetscObjectGetComm((PetscObject)ksp, &comm));
395: /* scatterv rhs */
396: PetscCallMPI(MPI_Comm_size(comm, &size));
397: if (pc) {
398: PetscInt N;
400: PCMPISolveCounts[size - 1]++;
401: PetscCall(MatGetSize(pc->pmat, &N, NULL));
402: PCMPISizes[size - 1] += N;
403: }
404: PetscCall(VecGetLocalSize(ksp->vec_rhs, &n));
405: PetscCall(PetscLogEventBegin(EventServerDistMPI, NULL, NULL, NULL, NULL));
406: if (!PCMPIServerUseShmget) {
407: PetscCall(VecGetArray(ksp->vec_rhs, &b));
408: if (pc) PetscCall(VecGetArrayRead(B, &sb));
409: PetscCallMPI(MPI_Scatterv(sb, pc ? km->sendcount : NULL, pc ? km->displ : NULL, MPIU_SCALAR, b, n, MPIU_SCALAR, 0, comm));
410: if (pc) PetscCall(VecRestoreArrayRead(B, &sb));
411: PetscCall(VecRestoreArray(ksp->vec_rhs, &b));
412: // TODO: scatter initial guess if needed
413: } else {
414: PetscInt rstart;
416: if (pc) PetscCall(VecGetArrayRead(B, &sb));
417: if (pc) PetscCall(VecGetArray(X, &sx));
418: const void *inaddr[2] = {(const void **)sb, (const void **)sx};
419: if (pc) PetscCall(VecRestoreArray(X, &sx));
420: if (pc) PetscCall(VecRestoreArrayRead(B, &sb));
422: PetscCall(PetscShmgetMapAddresses(comm, 2, inaddr, addr));
423: PetscCall(VecGetOwnershipRange(ksp->vec_rhs, &rstart, NULL));
424: PetscCall(VecPlaceArray(ksp->vec_rhs, rstart + (PetscScalar *)addr[0]));
425: PetscCall(VecPlaceArray(ksp->vec_sol, rstart + (PetscScalar *)addr[1]));
426: }
427: PetscCall(PetscLogEventEnd(EventServerDistMPI, NULL, NULL, NULL, NULL));
429: PetscCall(PetscLogEventEnd(EventServerDist, NULL, NULL, NULL, NULL));
430: PetscCall(PetscLogStagePush(PCMPIStage));
431: PetscCall(KSPSolve(ksp, NULL, NULL));
432: PetscCall(PetscLogStagePop());
433: PetscCall(PetscLogEventBegin(EventServerDist, NULL, NULL, NULL, NULL));
434: PetscCall(KSPGetIterationNumber(ksp, &its));
435: PCMPIIterations[size - 1] += its;
436: // TODO: send iterations up to outer KSP
438: if (PCMPIServerUseShmget) PetscCall(PetscShmgetUnmapAddresses(2, addr));
440: /* gather solution */
441: PetscCall(PetscLogEventBegin(EventServerDistMPI, NULL, NULL, NULL, NULL));
442: if (!PCMPIServerUseShmget) {
443: PetscCall(VecGetArrayRead(ksp->vec_sol, &x));
444: if (pc) PetscCall(VecGetArray(X, &sx));
445: PetscCallMPI(MPI_Gatherv(x, n, MPIU_SCALAR, sx, pc ? km->sendcount : NULL, pc ? km->displ : NULL, MPIU_SCALAR, 0, comm));
446: if (pc) PetscCall(VecRestoreArray(X, &sx));
447: PetscCall(VecRestoreArrayRead(ksp->vec_sol, &x));
448: } else {
449: PetscCall(VecResetArray(ksp->vec_rhs));
450: PetscCall(VecResetArray(ksp->vec_sol));
451: }
452: PetscCall(PetscLogEventEnd(EventServerDistMPI, NULL, NULL, NULL, NULL));
453: PetscCall(PetscLogEventEnd(EventServerDist, NULL, NULL, NULL, NULL));
454: PCMPIServerInSolve = PETSC_FALSE;
455: PetscFunctionReturn(PETSC_SUCCESS);
456: }
458: static PetscErrorCode PCMPIDestroy(PC pc)
459: {
460: PC_MPI *km = pc ? (PC_MPI *)pc->data : NULL;
461: KSP ksp;
462: MPI_Comm comm = PC_MPI_COMM_WORLD;
464: PetscFunctionBegin;
465: PetscCallMPI(MPI_Scatter(pc ? km->ksps : NULL, 1, MPI_AINT, &ksp, 1, MPI_AINT, 0, comm));
466: if (!ksp) PetscFunctionReturn(PETSC_SUCCESS);
467: PetscCall(PetscLogStagePush(PCMPIStage));
468: PCMPIServerInSolve = PETSC_TRUE;
469: PetscCall(KSPDestroy(&ksp));
470: PetscCall(PetscLogStagePop());
471: PCMPIServerInSolve = PETSC_FALSE;
472: PetscFunctionReturn(PETSC_SUCCESS);
473: }
475: static PetscErrorCode PCMPIServerBroadcastRequest(PCMPICommand request)
476: {
477: #if defined(PETSC_HAVE_PTHREAD_MUTEX)
478: PetscMPIInt dummy1 = 1, dummy2;
479: #endif
481: PetscFunctionBegin;
482: #if defined(PETSC_HAVE_PTHREAD_MUTEX)
483: if (PCMPIServerUseShmget) {
484: for (PetscMPIInt i = 1; i < PetscGlobalSize; i++) pthread_mutex_unlock(&PCMPIServerLocks[i]);
485: }
486: #endif
487: PetscCallMPI(MPI_Bcast(&request, 1, MPIU_ENUM, 0, MPI_COMM_WORLD));
488: /* next line ensures the sender has already taken the lock */
489: #if defined(PETSC_HAVE_PTHREAD_MUTEX)
490: if (PCMPIServerUseShmget) {
491: PetscCallMPI(MPI_Reduce(&dummy1, &dummy2, 1, MPI_INT, MPI_SUM, 0, PC_MPI_COMM_WORLD));
492: for (PetscMPIInt i = 1; i < PetscGlobalSize; i++) pthread_mutex_lock(&PCMPIServerLocks[i]);
493: }
494: #endif
495: PetscFunctionReturn(PETSC_SUCCESS);
496: }
498: /*@C
499: PCMPIServerBegin - starts a server that runs on the `rank != 0` MPI processes waiting to process requests for
500: parallel `KSP` solves and management of parallel `KSP` objects.
502: Logically Collective on all MPI processes except rank 0
504: Options Database Keys:
505: + -mpi_linear_solver_server - causes the PETSc program to start in MPI linear solver server mode where only the first MPI rank runs user code
506: . -mpi_linear_solver_server_view - displays information about all the linear systems solved by the MPI linear solver server at the conclusion of the program
507: - -mpi_linear_solver_server_use_shared_memory - use shared memory when communicating matrices and vectors to server processes (default where supported)
509: Level: developer
511: Note:
512: This is normally started automatically in `PetscInitialize()` when the option is provided
514: See `PCMPI` for information on using the solver with a `KSP` object
516: Developer Notes:
517: When called on MPI rank 0 this sets `PETSC_COMM_WORLD` to `PETSC_COMM_SELF` to allow a main program
518: written with `PETSC_COMM_WORLD` to run correctly on the single rank while all the ranks
519: (that would normally be sharing `PETSC_COMM_WORLD`) to run the solver server.
521: Can this be integrated into the `PetscDevice` abstraction that is currently being developed?
523: Conceivably `PCREDISTRIBUTE` could be organized in a similar manner to simplify its usage
525: This could be implemented directly at the `KSP` level instead of using the `PCMPI` wrapper object
527: The code could be extended to allow an MPI + OpenMP application to use the linear solver server concept across all shared-memory
528: nodes with a single MPI process per node for the user application but multiple MPI processes per node for the linear solver.
530: The concept could also be extended for users's callbacks for `SNES`, `TS`, and `Tao` where the `SNESSolve()` for example, runs on
531: all MPI processes but the user callback only runs on one MPI process per node.
533: PETSc could also be extended with an MPI-less API that provides access to PETSc's solvers without any reference to MPI, essentially remove
534: the `MPI_Comm` argument from PETSc calls.
536: .seealso: [](sec_pcmpi), `PCMPIServerEnd()`, `PCMPI`, `KSPCheckPCMPI()`
537: @*/
538: PetscErrorCode PCMPIServerBegin(void)
539: {
540: PetscMPIInt rank;
542: PetscFunctionBegin;
543: PetscCall(PetscInfo(NULL, "Starting MPI Linear Solver Server\n"));
544: if (PetscDefined(USE_SINGLE_LIBRARY)) {
545: PetscCall(VecInitializePackage());
546: PetscCall(MatInitializePackage());
547: PetscCall(DMInitializePackage());
548: PetscCall(PCInitializePackage());
549: PetscCall(KSPInitializePackage());
550: PetscCall(SNESInitializePackage());
551: PetscCall(TSInitializePackage());
552: PetscCall(TaoInitializePackage());
553: }
554: PetscCall(PetscLogStageRegister("PCMPI", &PCMPIStage));
555: PetscCall(PetscLogEventRegister("ServerDist", PC_CLASSID, &EventServerDist));
556: PetscCall(PetscLogEventRegister("ServerDistMPI", PC_CLASSID, &EventServerDistMPI));
558: if (!PetscDefined(HAVE_SHMGET)) PCMPIServerUseShmget = PETSC_FALSE;
559: PetscCall(PetscOptionsGetBool(NULL, NULL, "-mpi_linear_solver_server_use_shared_memory", &PCMPIServerUseShmget, NULL));
561: PetscCallMPI(MPI_Comm_rank(PC_MPI_COMM_WORLD, &rank));
562: if (PCMPIServerUseShmget) {
563: #if defined(PETSC_HAVE_PTHREAD_MUTEX)
564: PetscMPIInt size;
566: PetscCallMPI(MPI_Comm_size(PETSC_COMM_WORLD, &size));
567: if (size > 1) {
568: pthread_mutex_t *locks;
570: if (rank == 0) {
571: PCMPIServerActive = PETSC_TRUE;
572: PetscCall(PetscShmgetAllocateArray(size, sizeof(pthread_mutex_t), (void **)&locks));
573: }
574: PetscCall(PetscShmgetMapAddresses(PETSC_COMM_WORLD, 1, (const void **)&locks, (void **)&PCMPIServerLocks));
575: if (rank == 0) {
576: pthread_mutexattr_t attr;
578: pthread_mutexattr_init(&attr);
579: pthread_mutexattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
581: for (int i = 1; i < size; i++) {
582: pthread_mutex_init(&PCMPIServerLocks[i], &attr);
583: pthread_mutex_lock(&PCMPIServerLocks[i]);
584: }
585: }
586: PetscCallMPI(MPI_Barrier(PETSC_COMM_WORLD));
587: }
588: #endif
589: }
590: if (rank == 0) {
591: PETSC_COMM_WORLD = PETSC_COMM_SELF;
592: PCMPIServerActive = PETSC_TRUE;
593: PetscFunctionReturn(PETSC_SUCCESS);
594: }
596: while (PETSC_TRUE) {
597: PCMPICommand request = PCMPI_CREATE;
598: PetscMPIInt dummy1 = 1, dummy2;
600: // TODO: can we broadcast the number of active ranks here so only the correct subset of proccesses waits on the later scatters?
601: #if defined(PETSC_HAVE_PTHREAD_MUTEX)
602: if (PCMPIServerUseShmget) pthread_mutex_lock(&PCMPIServerLocks[PetscGlobalRank]);
603: #endif
604: PetscCallMPI(MPI_Bcast(&request, 1, MPIU_ENUM, 0, PC_MPI_COMM_WORLD));
605: #if defined(PETSC_HAVE_PTHREAD_MUTEX)
606: if (PCMPIServerUseShmget) {
607: /* next line ensures PetscGlobalRank has locked before rank 0 can take the lock back */
608: PetscCallMPI(MPI_Reduce(&dummy1, &dummy2, 1, MPI_INT, MPI_SUM, 0, PC_MPI_COMM_WORLD));
609: pthread_mutex_unlock(&PCMPIServerLocks[PetscGlobalRank]);
610: }
611: #endif
612: switch (request) {
613: case PCMPI_CREATE:
614: PetscCall(PCMPICreate(NULL));
615: break;
616: case PCMPI_SET_MAT:
617: PetscCall(PCMPISetMat(NULL));
618: break;
619: case PCMPI_UPDATE_MAT_VALUES:
620: PetscCall(PCMPIUpdateMatValues(NULL));
621: break;
622: case PCMPI_VIEW:
623: // PetscCall(PCMPIView(NULL));
624: break;
625: case PCMPI_SOLVE:
626: PetscCall(PCMPISolve(NULL, NULL, NULL));
627: break;
628: case PCMPI_DESTROY:
629: PetscCall(PCMPIDestroy(NULL));
630: break;
631: case PCMPI_EXIT:
632: if (PCMPIServerUseShmget) PetscCall(PetscShmgetUnmapAddresses(1, (void **)&PCMPIServerLocks));
633: PetscCall(PetscFinalize());
634: exit(0); /* not sure if this is a good idea, but cannot return because it will run users main program */
635: break;
636: default:
637: break;
638: }
639: }
640: PetscFunctionReturn(PETSC_SUCCESS);
641: }
643: /*@C
644: PCMPIServerEnd - ends a server that runs on the rank != 0 MPI processes waiting to process requests for
645: parallel KSP solves and management of parallel `KSP` objects.
647: Logically Collective on all MPI ranks except 0
649: Level: developer
651: Note:
652: This is normally called automatically in `PetscFinalize()`
654: .seealso: [](sec_pcmpi), `PCMPIServerBegin()`, `PCMPI`, `KSPCheckPCMPI()`
655: @*/
656: PetscErrorCode PCMPIServerEnd(void)
657: {
658: PetscFunctionBegin;
659: if (PetscGlobalRank == 0) {
660: PetscViewer viewer = NULL;
661: PetscViewerFormat format;
663: PetscCall(PetscShmgetAddressesFinalize());
664: PetscCall(PCMPIServerBroadcastRequest(PCMPI_EXIT));
665: if (PCMPIServerUseShmget) PetscCall(PetscShmgetUnmapAddresses(1, (void **)&PCMPIServerLocks));
666: PETSC_COMM_WORLD = MPI_COMM_WORLD; /* could use PC_MPI_COMM_WORLD */
667: PetscOptionsBegin(PETSC_COMM_SELF, NULL, "MPI linear solver server options", NULL);
668: PetscCall(PetscOptionsViewer("-mpi_linear_solver_server_view", "View information about system solved with the server", "PCMPI", &viewer, &format, NULL));
669: PetscOptionsEnd();
670: if (viewer) {
671: PetscBool isascii;
673: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
674: if (isascii) {
675: PetscMPIInt size;
676: PetscMPIInt i;
678: PetscCallMPI(MPI_Comm_size(PETSC_COMM_WORLD, &size));
679: PetscCall(PetscViewerASCIIPrintf(viewer, "MPI linear solver server statistics:\n"));
680: PetscCall(PetscViewerASCIIPrintf(viewer, " Ranks KSPSolve()s Mats KSPs Avg. Size Avg. Its\n"));
681: if (PCMPIKSPCountsSeq) {
682: PetscCall(PetscViewerASCIIPrintf(viewer, " Sequential %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT "\n", PCMPISolveCountsSeq, PCMPIKSPCountsSeq, PCMPISizesSeq / PCMPISolveCountsSeq, PCMPIIterationsSeq / PCMPISolveCountsSeq));
683: }
684: for (i = 0; i < size; i++) {
685: if (PCMPIKSPCounts[i]) {
686: PetscCall(PetscViewerASCIIPrintf(viewer, " %d %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT "\n", i + 1, PCMPISolveCounts[i], PCMPIMatCounts[i], PCMPIKSPCounts[i], PCMPISizes[i] / PCMPISolveCounts[i], PCMPIIterations[i] / PCMPISolveCounts[i]));
687: }
688: }
689: PetscCall(PetscViewerASCIIPrintf(viewer, "MPI linear solver server %susing shared memory\n", PCMPIServerUseShmget ? "" : "not "));
690: }
691: PetscCall(PetscViewerDestroy(&viewer));
692: }
693: }
694: PetscCall(PCMPICommsDestroy());
695: PCMPIServerActive = PETSC_FALSE;
696: PetscFunctionReturn(PETSC_SUCCESS);
697: }
699: /*
700: This version is used in the trivial case when the MPI parallel solver server is running on just the original MPI rank 0
701: because, for example, the problem is small. This version is more efficient because it does not require copying any data
702: */
703: static PetscErrorCode PCSetUp_Seq(PC pc)
704: {
705: PC_MPI *km = (PC_MPI *)pc->data;
706: Mat sA;
707: const char *prefix;
708: char *found = NULL, *cprefix;
710: PetscFunctionBegin;
711: PCMPIServerInSolve = PETSC_TRUE;
712: PetscCall(PCGetOperators(pc, NULL, &sA));
713: PetscCall(PCGetOptionsPrefix(pc, &prefix));
714: PetscCall(KSPCreate(PETSC_COMM_SELF, &km->ksps[0]));
715: PetscCall(KSPSetNestLevel(km->ksps[0], 1));
716: PetscCall(PetscObjectSetTabLevel((PetscObject)km->ksps[0], 1));
718: /* Created KSP gets prefix of PC minus the mpi_linear_solver_server_ portion */
719: PetscCall(PCGetOptionsPrefix(pc, &prefix));
720: PetscCheck(prefix, PETSC_COMM_SELF, PETSC_ERR_PLIB, "PCMPI missing required prefix");
721: PetscCall(PetscStrallocpy(prefix, &cprefix));
722: PetscCall(PetscStrstr(cprefix, "mpi_linear_solver_server_", &found));
723: PetscCheck(found, PETSC_COMM_SELF, PETSC_ERR_PLIB, "PCMPI missing mpi_linear_solver_server_ portion of prefix");
724: *found = 0;
725: PetscCall(KSPSetOptionsPrefix(km->ksps[0], cprefix));
726: PetscCall(PetscFree(cprefix));
728: PetscCall(KSPSetOperators(km->ksps[0], sA, sA));
729: PetscCall(KSPSetFromOptions(km->ksps[0]));
730: PetscCall(KSPSetUp(km->ksps[0]));
731: PetscCall(PetscInfo((PetscObject)pc, "MPI parallel linear solver system is being solved directly on rank 0 due to its small size\n"));
732: PCMPIKSPCountsSeq++;
733: PCMPIServerInSolve = PETSC_FALSE;
734: PetscFunctionReturn(PETSC_SUCCESS);
735: }
737: static PetscErrorCode PCApply_Seq(PC pc, Vec b, Vec x)
738: {
739: PC_MPI *km = (PC_MPI *)pc->data;
740: PetscInt its, n;
741: Mat A;
743: PetscFunctionBegin;
744: PCMPIServerInSolve = PETSC_TRUE;
745: PetscCall(KSPSolve(km->ksps[0], b, x));
746: PetscCall(KSPGetIterationNumber(km->ksps[0], &its));
747: PCMPISolveCountsSeq++;
748: PCMPIIterationsSeq += its;
749: PetscCall(KSPGetOperators(km->ksps[0], NULL, &A));
750: PetscCall(MatGetSize(A, &n, NULL));
751: PCMPISizesSeq += n;
752: PCMPIServerInSolve = PETSC_FALSE;
753: /*
754: do not keep reference to previous rhs and solution since destroying them in the next KSPSolve()
755: my use PetscFree() instead of PCMPIArrayDeallocate()
756: */
757: PetscCall(VecDestroy(&km->ksps[0]->vec_rhs));
758: PetscCall(VecDestroy(&km->ksps[0]->vec_sol));
759: PetscFunctionReturn(PETSC_SUCCESS);
760: }
762: static PetscErrorCode PCView_Seq(PC pc, PetscViewer viewer)
763: {
764: PC_MPI *km = (PC_MPI *)pc->data;
766: PetscFunctionBegin;
767: PetscCall(PetscViewerASCIIPrintf(viewer, "Running MPI linear solver server directly on rank 0 due to its small size\n"));
768: PetscCall(PetscViewerASCIIPrintf(viewer, "Desired minimum number of nonzeros per rank for MPI parallel solve %d\n", (int)km->mincntperrank));
769: PetscCall(PetscViewerASCIIPrintf(viewer, "*** Use -mpi_linear_solver_server_view to statistics on all the solves ***\n"));
770: PetscFunctionReturn(PETSC_SUCCESS);
771: }
773: static PetscErrorCode PCDestroy_Seq(PC pc)
774: {
775: PC_MPI *km = (PC_MPI *)pc->data;
776: Mat A, B;
777: Vec x, b;
779: PetscFunctionBegin;
780: PCMPIServerInSolve = PETSC_TRUE;
781: /* since matrices and vectors are shared with outer KSP we need to ensure they are not destroyed with PetscFree() */
782: PetscCall(KSPGetOperators(km->ksps[0], &A, &B));
783: PetscCall(PetscObjectReference((PetscObject)A));
784: PetscCall(PetscObjectReference((PetscObject)B));
785: PetscCall(KSPGetSolution(km->ksps[0], &x));
786: PetscCall(PetscObjectReference((PetscObject)x));
787: PetscCall(KSPGetRhs(km->ksps[0], &b));
788: PetscCall(PetscObjectReference((PetscObject)b));
789: PetscCall(KSPDestroy(&km->ksps[0]));
790: PetscCall(PetscFree(pc->data));
791: PCMPIServerInSolve = PETSC_FALSE;
792: PetscCall(MatDestroy(&A));
793: PetscCall(MatDestroy(&B));
794: PetscCall(VecDestroy(&x));
795: PetscCall(VecDestroy(&b));
796: PetscFunctionReturn(PETSC_SUCCESS);
797: }
799: /*
800: PCSetUp_MPI - Trigger the creation of the MPI parallel PC and copy parts of the matrix and
801: right-hand side to the parallel PC
802: */
803: static PetscErrorCode PCSetUp_MPI(PC pc)
804: {
805: PC_MPI *km = (PC_MPI *)pc->data;
806: PetscMPIInt rank, size;
807: PetscBool newmatrix = PETSC_FALSE;
809: PetscFunctionBegin;
810: PetscCallMPI(MPI_Comm_rank(MPI_COMM_WORLD, &rank));
811: PetscCheck(rank == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "PCMPI can only be used from 0th rank of MPI_COMM_WORLD. Perhaps a missing -mpi_linear_solver_server?");
812: PetscCallMPI(MPI_Comm_size(MPI_COMM_WORLD, &size));
814: if (!pc->setupcalled) {
815: if (!km->alwaysuseserver) {
816: PetscInt n;
817: Mat sA;
818: /* short circuit for small systems */
819: PetscCall(PCGetOperators(pc, &sA, &sA));
820: PetscCall(MatGetSize(sA, &n, NULL));
821: if (n < 2 * km->mincntperrank - 1 || size == 1) {
822: pc->ops->setup = NULL;
823: pc->ops->apply = PCApply_Seq;
824: pc->ops->destroy = PCDestroy_Seq;
825: pc->ops->view = PCView_Seq;
826: PetscCall(PCSetUp_Seq(pc));
827: PetscFunctionReturn(PETSC_SUCCESS);
828: }
829: }
831: PetscCall(PCMPIServerBroadcastRequest(PCMPI_CREATE));
832: PetscCall(PCMPICreate(pc));
833: newmatrix = PETSC_TRUE;
834: }
835: if (pc->flag == DIFFERENT_NONZERO_PATTERN) newmatrix = PETSC_TRUE;
837: if (newmatrix) {
838: PetscCall(PetscInfo((PetscObject)pc, "New matrix or matrix has changed nonzero structure\n"));
839: PetscCall(PCMPIServerBroadcastRequest(PCMPI_SET_MAT));
840: PetscCall(PCMPISetMat(pc));
841: } else {
842: PetscCall(PetscInfo((PetscObject)pc, "Matrix has only changed nonzero values\n"));
843: PetscCall(PCMPIServerBroadcastRequest(PCMPI_UPDATE_MAT_VALUES));
844: PetscCall(PCMPIUpdateMatValues(pc));
845: }
846: PetscFunctionReturn(PETSC_SUCCESS);
847: }
849: static PetscErrorCode PCApply_MPI(PC pc, Vec b, Vec x)
850: {
851: PetscFunctionBegin;
852: PetscCall(PCMPIServerBroadcastRequest(PCMPI_SOLVE));
853: PetscCall(PCMPISolve(pc, b, x));
854: PetscFunctionReturn(PETSC_SUCCESS);
855: }
857: static PetscErrorCode PCDestroy_MPI(PC pc)
858: {
859: PetscFunctionBegin;
860: PetscCall(PCMPIServerBroadcastRequest(PCMPI_DESTROY));
861: PetscCall(PCMPIDestroy(pc));
862: PetscCall(PetscFree(pc->data));
863: PetscFunctionReturn(PETSC_SUCCESS);
864: }
866: /*
867: PCView_MPI - Cannot call view on the MPI parallel KSP because other ranks do not have access to the viewer, use options database
868: */
869: static PetscErrorCode PCView_MPI(PC pc, PetscViewer viewer)
870: {
871: PC_MPI *km = (PC_MPI *)pc->data;
872: MPI_Comm comm;
873: PetscMPIInt size;
875: PetscFunctionBegin;
876: PetscCall(PetscObjectGetComm((PetscObject)km->ksps[0], &comm));
877: PetscCallMPI(MPI_Comm_size(comm, &size));
878: PetscCall(PetscViewerASCIIPrintf(viewer, "Size of MPI communicator used for MPI parallel KSP solve %d\n", size));
879: PetscCall(PetscViewerASCIIPrintf(viewer, "Desired minimum number of matrix rows on each MPI process for MPI parallel solve %d\n", (int)km->mincntperrank));
880: PetscCall(PetscViewerASCIIPrintf(viewer, "*** Use -mpi_linear_solver_server_view to view statistics on all the solves ***\n"));
881: PetscFunctionReturn(PETSC_SUCCESS);
882: }
884: static PetscErrorCode PCSetFromOptions_MPI(PC pc, PetscOptionItems *PetscOptionsObject)
885: {
886: PC_MPI *km = (PC_MPI *)pc->data;
888: PetscFunctionBegin;
889: PetscOptionsHeadBegin(PetscOptionsObject, "MPI linear solver server options");
890: PetscCall(PetscOptionsInt("-minimum_count_per_rank", "Desired minimum number of nonzeros per rank", "None", km->mincntperrank, &km->mincntperrank, NULL));
891: PetscCall(PetscOptionsBool("-always_use_server", "Use the server even if only one rank is used for the solve (for debugging)", "None", km->alwaysuseserver, &km->alwaysuseserver, NULL));
892: PetscOptionsHeadEnd();
893: PetscFunctionReturn(PETSC_SUCCESS);
894: }
896: /*MC
897: PCMPI - Calls an MPI parallel `KSP` to solve a linear system from user code running on one process
899: Options Database Keys for the Server:
900: + -mpi_linear_solver_server - causes the PETSc program to start in MPI linear solver server mode where only the first MPI rank runs user code
901: . -mpi_linear_solver_server_view - displays information about all the linear systems solved by the MPI linear solver server
902: - -mpi_linear_solver_server_use_shared_memory <true, false> - use shared memory to distribute matrix and right hand side, defaults to true
904: Options Database Keys for a specific `KSP` object
905: + -[any_ksp_prefix]_mpi_linear_solver_server_minimum_count_per_rank - sets the minimum size of the linear system per MPI rank that the solver will strive for
906: - -[any_ksp_prefix]_mpi_linear_solver_server_always_use_server - use the server solver code even if the particular system is only solved on the process (for debugging and testing purposes)
908: Level: developer
910: Notes:
911: This cannot be used with vectors or matrices that are created using arrays provided by the user, such as `VecCreateWithArray()` or
912: `MatCreateSeqAIJWithArrays()`
914: The options database prefix for the actual solver is any prefix provided before use to the original `KSP` with `KSPSetOptionsPrefix()`, mostly commonly no prefix is used.
916: It can be particularly useful for user OpenMP code or potentially user GPU code.
918: When the program is running with a single MPI process then it directly uses the provided matrix and right-hand side
919: and does not need to distribute the matrix and vector to the various MPI processes; thus it incurs no extra overhead over just using the `KSP` directly.
921: The solver options for actual solving `KSP` and `PC` must be controlled via the options database, calls to set options directly on the user level `KSP` and `PC` have no effect
922: because they are not the actual solver objects.
924: When `-log_view` is used with this solver the events within the parallel solve are logging in their own stage. Some of the logging in the other
925: stages will be confusing since the event times are only recorded on the 0th MPI rank, thus the percent of time in the events will be misleading.
927: Developer Note:
928: This `PCType` is never directly selected by the user, it is set when the option `-mpi_linear_solver_server` is used and the `PC` is at the outer most nesting of
929: a `KSP`. The outer most `KSP` object is automatically set to `KSPPREONLY` and thus is not directly visible to the user.
931: .seealso: [](sec_pcmpi), `KSPCreate()`, `KSPSetType()`, `KSPType`, `KSP`, `PC`, `PCMPIServerBegin()`, `PCMPIServerEnd()`, `KSPCheckPCMPI()`
932: M*/
933: PETSC_EXTERN PetscErrorCode PCCreate_MPI(PC pc)
934: {
935: PC_MPI *km;
936: char *found = NULL;
938: PetscFunctionBegin;
939: PetscCall(PetscStrstr(((PetscObject)pc)->prefix, "mpi_linear_solver_server_", &found));
940: PetscCheck(found, PETSC_COMM_SELF, PETSC_ERR_PLIB, "PCMPI object prefix does not have mpi_linear_solver_server_");
942: /* material from PCSetType() */
943: PetscTryTypeMethod(pc, destroy);
944: pc->ops->destroy = NULL;
945: pc->data = NULL;
947: PetscCall(PetscFunctionListDestroy(&((PetscObject)pc)->qlist));
948: PetscCall(PetscMemzero(pc->ops, sizeof(struct _PCOps)));
949: pc->modifysubmatrices = NULL;
950: pc->modifysubmatricesP = NULL;
951: pc->setupcalled = 0;
953: PetscCall(PetscNew(&km));
954: pc->data = (void *)km;
956: km->mincntperrank = 10000;
958: pc->ops->setup = PCSetUp_MPI;
959: pc->ops->apply = PCApply_MPI;
960: pc->ops->destroy = PCDestroy_MPI;
961: pc->ops->view = PCView_MPI;
962: pc->ops->setfromoptions = PCSetFromOptions_MPI;
963: PetscCall(PetscObjectChangeTypeName((PetscObject)pc, PCMPI));
964: PetscFunctionReturn(PETSC_SUCCESS);
965: }
967: /*@
968: PCMPIGetKSP - Gets the `KSP` created by the `PCMPI`
970: Not Collective
972: Input Parameter:
973: . pc - the preconditioner context
975: Output Parameter:
976: . innerksp - the inner `KSP`
978: Level: advanced
980: .seealso: [](ch_ksp), `KSP`, `PCMPI`, `PCREDISTRIBUTE`
981: @*/
982: PetscErrorCode PCMPIGetKSP(PC pc, KSP *innerksp)
983: {
984: PC_MPI *red = (PC_MPI *)pc->data;
986: PetscFunctionBegin;
988: PetscAssertPointer(innerksp, 2);
989: *innerksp = red->ksps[0];
990: PetscFunctionReturn(PETSC_SUCCESS);
991: }