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: PetscInt sendcount[PC_MPI_MAX_RANKS], displ[PC_MPI_MAX_RANKS]; /* For scatter/gather of rhs/solution */
27: PetscInt 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: PetscCall(PetscMPIIntCast(slen, &len));
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: PetscInt *NZ = NULL, sendcounti[PC_MPI_MAX_RANKS], displi[PC_MPI_MAX_RANKS], *NZdispl = NULL, nz;
168: PetscMPIInt size, i;
169: const PetscScalar *a = NULL, *sa;
170: PetscInt matproperties[8] = {0}, rstart, rend;
171: char *cprefix;
173: PetscFunctionBegin;
174: PetscCallMPI(MPI_Scatter(pc ? km->ksps : NULL, 1, MPI_AINT, &ksp, 1, MPI_AINT, 0, comm));
175: if (!ksp) PetscFunctionReturn(PETSC_SUCCESS);
176: PCMPIServerInSolve = PETSC_TRUE;
177: PetscCall(PetscLogEventBegin(EventServerDist, NULL, NULL, NULL, NULL));
178: PetscCall(PetscObjectGetComm((PetscObject)ksp, &comm));
179: if (pc) {
180: PetscBool isset, issymmetric, ishermitian, isspd, isstructurallysymmetric;
181: const char *prefix;
182: size_t clen;
184: PetscCallMPI(MPI_Comm_size(comm, &size));
185: PCMPIMatCounts[size - 1]++;
186: PetscCall(PCGetOperators(pc, &sA, &sA));
187: PetscCall(MatGetSize(sA, &matproperties[0], &matproperties[1]));
188: PetscCall(MatGetBlockSize(sA, &bs));
189: matproperties[2] = bs;
190: PetscCall(MatIsSymmetricKnown(sA, &isset, &issymmetric));
191: matproperties[3] = !isset ? 0 : (issymmetric ? 1 : 2);
192: PetscCall(MatIsHermitianKnown(sA, &isset, &ishermitian));
193: matproperties[4] = !isset ? 0 : (ishermitian ? 1 : 2);
194: PetscCall(MatIsSPDKnown(sA, &isset, &isspd));
195: matproperties[5] = !isset ? 0 : (isspd ? 1 : 2);
196: PetscCall(MatIsStructurallySymmetricKnown(sA, &isset, &isstructurallysymmetric));
197: matproperties[6] = !isset ? 0 : (isstructurallysymmetric ? 1 : 2);
198: /* Created Mat gets prefix of input Mat PLUS the mpi_linear_solver_server_ portion */
199: PetscCall(MatGetOptionsPrefix(sA, &prefix));
200: PetscCall(PetscStrallocpy(prefix, &cprefix));
201: PetscCall(PetscStrlen(cprefix, &clen));
202: matproperties[7] = (PetscInt)clen;
203: }
204: PetscCallMPI(MPI_Bcast(matproperties, PETSC_STATIC_ARRAY_LENGTH(matproperties), MPIU_INT, 0, comm));
206: /* determine ownership ranges of matrix columns */
207: PetscCall(PetscLayoutCreate(comm, &layout));
208: PetscCall(PetscLayoutSetBlockSize(layout, matproperties[2]));
209: PetscCall(PetscLayoutSetSize(layout, matproperties[1]));
210: PetscCall(PetscLayoutSetUp(layout));
211: PetscCall(PetscLayoutGetLocalSize(layout, &n));
212: PetscCall(PetscLayoutDestroy(&layout));
214: /* determine ownership ranges of matrix rows */
215: PetscCall(PetscLayoutCreate(comm, &layout));
216: PetscCall(PetscLayoutSetBlockSize(layout, matproperties[2]));
217: PetscCall(PetscLayoutSetSize(layout, matproperties[0]));
218: PetscCall(PetscLayoutSetUp(layout));
219: PetscCall(PetscLayoutGetLocalSize(layout, &m));
220: PetscCall(PetscLayoutGetRange(layout, &rstart, &rend));
222: PetscCall(PetscLogEventBegin(EventServerDistMPI, NULL, NULL, NULL, NULL));
223: /* copy over the matrix nonzero structure and values */
224: if (pc) {
225: PetscCall(MatGetRowIJ(sA, 0, PETSC_FALSE, PETSC_FALSE, NULL, &IA, &JA, NULL));
226: if (!PCMPIServerUseShmget) {
227: NZ = km->NZ;
228: NZdispl = km->NZdispl;
229: PetscCall(PetscLayoutGetRanges(layout, &range));
230: for (i = 0; i < size; i++) {
231: sendcounti[i] = 1 + range[i + 1] - range[i];
232: NZ[i] = IA[range[i + 1]] - IA[range[i]];
233: }
234: displi[0] = 0;
235: NZdispl[0] = 0;
236: for (j = 1; j < size; j++) {
237: displi[j] = displi[j - 1] + sendcounti[j - 1] - 1;
238: NZdispl[j] = NZdispl[j - 1] + NZ[j - 1];
239: }
240: }
241: PetscCall(MatSeqAIJGetArrayRead(sA, &sa));
242: }
243: PetscCall(PetscLayoutDestroy(&layout));
245: PetscCall(MatCreate(comm, &A));
246: if (matproperties[7] > 0) {
247: PetscMPIInt ni;
249: PetscCall(PetscMPIIntCast(matproperties[7] + 1, &ni));
250: if (!pc) PetscCall(PetscMalloc1(matproperties[7] + 1, &cprefix));
251: PetscCallMPI(MPI_Bcast(cprefix, ni, MPI_CHAR, 0, comm));
252: PetscCall(MatSetOptionsPrefix(A, cprefix));
253: PetscCall(PetscFree(cprefix));
254: }
255: PetscCall(MatAppendOptionsPrefix(A, "mpi_linear_solver_server_"));
256: PetscCall(MatSetSizes(A, m, n, matproperties[0], matproperties[1]));
257: PetscCall(MatSetType(A, MATMPIAIJ));
259: if (!PCMPIServerUseShmget) {
260: PetscCallMPI(MPI_Scatter(NZ, 1, MPIU_INT, &nz, 1, MPIU_INT, 0, comm));
261: PetscCall(PetscMalloc3(n + 1, &ia, nz, &ja, nz, &a));
262: PetscCallMPI(MPIU_Scatterv(IA, sendcounti, displi, MPIU_INT, (void *)ia, n + 1, MPIU_INT, 0, comm));
263: PetscCallMPI(MPIU_Scatterv(JA, NZ, NZdispl, MPIU_INT, (void *)ja, nz, MPIU_INT, 0, comm));
264: PetscCallMPI(MPIU_Scatterv(sa, NZ, NZdispl, MPIU_SCALAR, (void *)a, nz, MPIU_SCALAR, 0, comm));
265: } else {
266: const void *addr[3] = {(const void **)IA, (const void **)JA, (const void **)sa};
267: PCMPIServerAddresses *addresses;
269: PetscCall(PetscNew(&addresses));
270: addresses->n = 3;
271: PetscCall(PetscShmgetMapAddresses(comm, addresses->n, addr, addresses->addr));
272: ia = rstart + (PetscInt *)addresses->addr[0];
273: ja = ia[0] + (PetscInt *)addresses->addr[1];
274: a = ia[0] + (PetscScalar *)addresses->addr[2];
275: PetscCall(PetscObjectContainerCompose((PetscObject)A, "PCMPIServerAddresses", (void *)addresses, PCMPIServerAddressesDestroy));
276: }
278: if (pc) {
279: PetscCall(MatSeqAIJRestoreArrayRead(sA, &sa));
280: PetscCall(MatRestoreRowIJ(sA, 0, PETSC_FALSE, PETSC_FALSE, NULL, &IA, &JA, NULL));
281: }
282: PetscCall(PetscLogEventEnd(EventServerDistMPI, NULL, NULL, NULL, NULL));
284: PetscCall(PetscLogStagePush(PCMPIStage));
285: PetscCall(MatMPIAIJSetPreallocationCSR(A, ia, ja, a));
286: PetscCall(MatSetBlockSize(A, matproperties[2]));
288: if (matproperties[3]) PetscCall(MatSetOption(A, MAT_SYMMETRIC, matproperties[3] == 1 ? PETSC_TRUE : PETSC_FALSE));
289: if (matproperties[4]) PetscCall(MatSetOption(A, MAT_HERMITIAN, matproperties[4] == 1 ? PETSC_TRUE : PETSC_FALSE));
290: if (matproperties[5]) PetscCall(MatSetOption(A, MAT_SPD, matproperties[5] == 1 ? PETSC_TRUE : PETSC_FALSE));
291: if (matproperties[6]) PetscCall(MatSetOption(A, MAT_STRUCTURALLY_SYMMETRIC, matproperties[6] == 1 ? PETSC_TRUE : PETSC_FALSE));
293: if (!PCMPIServerUseShmget) PetscCall(PetscFree3(ia, ja, a));
294: PetscCall(KSPSetOperators(ksp, A, A));
295: if (!ksp->vec_sol) PetscCall(MatCreateVecs(A, &ksp->vec_sol, &ksp->vec_rhs));
296: PetscCall(PetscLogStagePop());
297: if (pc && !PCMPIServerUseShmget) { /* needed for scatterv/gatherv of rhs and solution */
298: const PetscInt *range;
300: PetscCall(VecGetOwnershipRanges(ksp->vec_sol, &range));
301: for (i = 0; i < size; i++) {
302: km->sendcount[i] = range[i + 1] - range[i];
303: km->displ[i] = range[i];
304: }
305: }
306: PetscCall(MatDestroy(&A));
307: PetscCall(PetscLogEventEnd(EventServerDist, NULL, NULL, NULL, NULL));
308: PetscCall(KSPSetFromOptions(ksp));
309: PCMPIServerInSolve = PETSC_FALSE;
310: PetscFunctionReturn(PETSC_SUCCESS);
311: }
313: static PetscErrorCode PCMPIUpdateMatValues(PC pc)
314: {
315: PC_MPI *km = pc ? (PC_MPI *)pc->data : NULL;
316: KSP ksp;
317: Mat sA, A;
318: MPI_Comm comm = PC_MPI_COMM_WORLD;
319: const PetscInt *ia, *IA;
320: const PetscScalar *a;
321: PetscCount nz;
322: const PetscScalar *sa = NULL;
323: PetscMPIInt size;
324: PetscInt rstart, matproperties[4] = {0, 0, 0, 0};
326: PetscFunctionBegin;
327: if (pc) {
328: PetscCall(PCGetOperators(pc, &sA, &sA));
329: PetscCall(MatSeqAIJGetArrayRead(sA, &sa));
330: PetscCall(MatGetRowIJ(sA, 0, PETSC_FALSE, PETSC_FALSE, NULL, &IA, NULL, NULL));
331: }
332: PetscCallMPI(MPI_Scatter(pc ? km->ksps : NULL, 1, MPI_AINT, &ksp, 1, MPI_AINT, 0, comm));
333: if (!ksp) PetscFunctionReturn(PETSC_SUCCESS);
334: PCMPIServerInSolve = PETSC_TRUE;
335: PetscCall(PetscLogEventBegin(EventServerDist, NULL, NULL, NULL, NULL));
336: PetscCall(PetscObjectGetComm((PetscObject)ksp, &comm));
337: PetscCallMPI(MPI_Comm_size(comm, &size));
338: PCMPIMatCounts[size - 1]++;
339: PetscCall(KSPGetOperators(ksp, NULL, &A));
340: PetscCall(PetscLogEventBegin(EventServerDistMPI, NULL, NULL, NULL, NULL));
341: if (!PCMPIServerUseShmget) {
342: PetscInt petsc_nz;
344: PetscCall(MatMPIAIJGetNumberNonzeros(A, &nz));
345: PetscCall(PetscIntCast(nz, &petsc_nz));
346: PetscCall(PetscMalloc1(nz, &a));
347: PetscCallMPI(MPIU_Scatterv(sa, pc ? km->NZ : NULL, pc ? km->NZdispl : NULL, MPIU_SCALAR, (void *)a, petsc_nz, MPIU_SCALAR, 0, comm));
348: } else {
349: PetscCall(MatGetOwnershipRange(A, &rstart, NULL));
350: PCMPIServerAddresses *addresses;
351: PetscCall(PetscObjectContainerQuery((PetscObject)A, "PCMPIServerAddresses", (void **)&addresses));
352: ia = rstart + (PetscInt *)addresses->addr[0];
353: a = ia[0] + (PetscScalar *)addresses->addr[2];
354: }
355: PetscCall(PetscLogEventEnd(EventServerDistMPI, NULL, NULL, NULL, NULL));
356: if (pc) {
357: PetscBool isset, issymmetric, ishermitian, isspd, isstructurallysymmetric;
359: PetscCall(MatSeqAIJRestoreArrayRead(sA, &sa));
360: PetscCall(MatRestoreRowIJ(sA, 0, PETSC_FALSE, PETSC_FALSE, NULL, &IA, NULL, NULL));
362: PetscCall(MatIsSymmetricKnown(sA, &isset, &issymmetric));
363: matproperties[0] = !isset ? 0 : (issymmetric ? 1 : 2);
364: PetscCall(MatIsHermitianKnown(sA, &isset, &ishermitian));
365: matproperties[1] = !isset ? 0 : (ishermitian ? 1 : 2);
366: PetscCall(MatIsSPDKnown(sA, &isset, &isspd));
367: matproperties[2] = !isset ? 0 : (isspd ? 1 : 2);
368: PetscCall(MatIsStructurallySymmetricKnown(sA, &isset, &isstructurallysymmetric));
369: matproperties[3] = !isset ? 0 : (isstructurallysymmetric ? 1 : 2);
370: }
371: PetscCall(MatUpdateMPIAIJWithArray(A, a));
372: if (!PCMPIServerUseShmget) PetscCall(PetscFree(a));
373: PetscCallMPI(MPI_Bcast(matproperties, 4, MPIU_INT, 0, comm));
374: /* 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 */
375: if (matproperties[0]) PetscCall(MatSetOption(A, MAT_SYMMETRIC, matproperties[0] == 1 ? PETSC_TRUE : PETSC_FALSE));
376: if (matproperties[1]) PetscCall(MatSetOption(A, MAT_HERMITIAN, matproperties[1] == 1 ? PETSC_TRUE : PETSC_FALSE));
377: if (matproperties[2]) PetscCall(MatSetOption(A, MAT_SPD, matproperties[2] == 1 ? PETSC_TRUE : PETSC_FALSE));
378: if (matproperties[3]) PetscCall(MatSetOption(A, MAT_STRUCTURALLY_SYMMETRIC, matproperties[3] == 1 ? PETSC_TRUE : PETSC_FALSE));
379: PetscCall(PetscLogEventEnd(EventServerDist, NULL, NULL, NULL, NULL));
380: PCMPIServerInSolve = PETSC_FALSE;
381: PetscFunctionReturn(PETSC_SUCCESS);
382: }
384: static PetscErrorCode PCMPISolve(PC pc, Vec B, Vec X)
385: {
386: PC_MPI *km = pc ? (PC_MPI *)pc->data : NULL;
387: KSP ksp;
388: MPI_Comm comm = PC_MPI_COMM_WORLD;
389: const PetscScalar *sb = NULL, *x;
390: PetscScalar *b, *sx = NULL;
391: PetscInt its, n;
392: PetscMPIInt size;
393: void *addr[2];
395: PetscFunctionBegin;
396: PetscCallMPI(MPI_Scatter(pc ? km->ksps : &ksp, 1, MPI_AINT, &ksp, 1, MPI_AINT, 0, comm));
397: if (!ksp) PetscFunctionReturn(PETSC_SUCCESS);
398: PCMPIServerInSolve = PETSC_TRUE;
399: PetscCall(PetscLogEventBegin(EventServerDist, NULL, NULL, NULL, NULL));
400: PetscCall(PetscObjectGetComm((PetscObject)ksp, &comm));
402: /* scatterv rhs */
403: PetscCallMPI(MPI_Comm_size(comm, &size));
404: if (pc) {
405: PetscInt N;
407: PCMPISolveCounts[size - 1]++;
408: PetscCall(MatGetSize(pc->pmat, &N, NULL));
409: PCMPISizes[size - 1] += N;
410: }
411: PetscCall(VecGetLocalSize(ksp->vec_rhs, &n));
412: PetscCall(PetscLogEventBegin(EventServerDistMPI, NULL, NULL, NULL, NULL));
413: if (!PCMPIServerUseShmget) {
414: PetscCall(VecGetArray(ksp->vec_rhs, &b));
415: if (pc) PetscCall(VecGetArrayRead(B, &sb));
416: PetscCallMPI(MPIU_Scatterv(sb, pc ? km->sendcount : NULL, pc ? km->displ : NULL, MPIU_SCALAR, b, n, MPIU_SCALAR, 0, comm));
417: if (pc) PetscCall(VecRestoreArrayRead(B, &sb));
418: PetscCall(VecRestoreArray(ksp->vec_rhs, &b));
419: // TODO: scatter initial guess if needed
420: } else {
421: PetscInt rstart;
423: if (pc) PetscCall(VecGetArrayRead(B, &sb));
424: if (pc) PetscCall(VecGetArray(X, &sx));
425: const void *inaddr[2] = {(const void **)sb, (const void **)sx};
426: if (pc) PetscCall(VecRestoreArray(X, &sx));
427: if (pc) PetscCall(VecRestoreArrayRead(B, &sb));
429: PetscCall(PetscShmgetMapAddresses(comm, 2, inaddr, addr));
430: PetscCall(VecGetOwnershipRange(ksp->vec_rhs, &rstart, NULL));
431: PetscCall(VecPlaceArray(ksp->vec_rhs, rstart + (PetscScalar *)addr[0]));
432: PetscCall(VecPlaceArray(ksp->vec_sol, rstart + (PetscScalar *)addr[1]));
433: }
434: PetscCall(PetscLogEventEnd(EventServerDistMPI, NULL, NULL, NULL, NULL));
436: PetscCall(PetscLogEventEnd(EventServerDist, NULL, NULL, NULL, NULL));
437: PetscCall(PetscLogStagePush(PCMPIStage));
438: PetscCall(KSPSolve(ksp, NULL, NULL));
439: PetscCall(PetscLogStagePop());
440: PetscCall(PetscLogEventBegin(EventServerDist, NULL, NULL, NULL, NULL));
441: PetscCall(KSPGetIterationNumber(ksp, &its));
442: PCMPIIterations[size - 1] += its;
443: // TODO: send iterations up to outer KSP
445: if (PCMPIServerUseShmget) PetscCall(PetscShmgetUnmapAddresses(2, addr));
447: /* gather solution */
448: PetscCall(PetscLogEventBegin(EventServerDistMPI, NULL, NULL, NULL, NULL));
449: if (!PCMPIServerUseShmget) {
450: PetscCall(VecGetArrayRead(ksp->vec_sol, &x));
451: if (pc) PetscCall(VecGetArray(X, &sx));
452: PetscCallMPI(MPIU_Gatherv(x, n, MPIU_SCALAR, sx, pc ? km->sendcount : NULL, pc ? km->displ : NULL, MPIU_SCALAR, 0, comm));
453: if (pc) PetscCall(VecRestoreArray(X, &sx));
454: PetscCall(VecRestoreArrayRead(ksp->vec_sol, &x));
455: } else {
456: PetscCall(VecResetArray(ksp->vec_rhs));
457: PetscCall(VecResetArray(ksp->vec_sol));
458: }
459: PetscCall(PetscLogEventEnd(EventServerDistMPI, NULL, NULL, NULL, NULL));
460: PetscCall(PetscLogEventEnd(EventServerDist, NULL, NULL, NULL, NULL));
461: PCMPIServerInSolve = PETSC_FALSE;
462: PetscFunctionReturn(PETSC_SUCCESS);
463: }
465: static PetscErrorCode PCMPIDestroy(PC pc)
466: {
467: PC_MPI *km = pc ? (PC_MPI *)pc->data : NULL;
468: KSP ksp;
469: MPI_Comm comm = PC_MPI_COMM_WORLD;
471: PetscFunctionBegin;
472: PetscCallMPI(MPI_Scatter(pc ? km->ksps : NULL, 1, MPI_AINT, &ksp, 1, MPI_AINT, 0, comm));
473: if (!ksp) PetscFunctionReturn(PETSC_SUCCESS);
474: PetscCall(PetscLogStagePush(PCMPIStage));
475: PCMPIServerInSolve = PETSC_TRUE;
476: PetscCall(KSPDestroy(&ksp));
477: PetscCall(PetscLogStagePop());
478: PCMPIServerInSolve = PETSC_FALSE;
479: PetscFunctionReturn(PETSC_SUCCESS);
480: }
482: static PetscErrorCode PCMPIServerBroadcastRequest(PCMPICommand request)
483: {
484: #if defined(PETSC_HAVE_PTHREAD_MUTEX)
485: PetscMPIInt dummy1 = 1, dummy2;
486: #endif
488: PetscFunctionBegin;
489: #if defined(PETSC_HAVE_PTHREAD_MUTEX)
490: if (PCMPIServerUseShmget) {
491: for (PetscMPIInt i = 1; i < PetscGlobalSize; i++) pthread_mutex_unlock(&PCMPIServerLocks[i]);
492: }
493: #endif
494: PetscCallMPI(MPI_Bcast(&request, 1, MPIU_ENUM, 0, MPI_COMM_WORLD));
495: /* next line ensures the sender has already taken the lock */
496: #if defined(PETSC_HAVE_PTHREAD_MUTEX)
497: if (PCMPIServerUseShmget) {
498: PetscCallMPI(MPI_Reduce(&dummy1, &dummy2, 1, MPI_INT, MPI_SUM, 0, PC_MPI_COMM_WORLD));
499: for (PetscMPIInt i = 1; i < PetscGlobalSize; i++) pthread_mutex_lock(&PCMPIServerLocks[i]);
500: }
501: #endif
502: PetscFunctionReturn(PETSC_SUCCESS);
503: }
505: /*@C
506: PCMPIServerBegin - starts a server that runs on the `rank != 0` MPI processes waiting to process requests for
507: parallel `KSP` solves and management of parallel `KSP` objects.
509: Logically Collective on all MPI processes except rank 0
511: Options Database Keys:
512: + -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
513: . -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
514: - -mpi_linear_solver_server_use_shared_memory - use shared memory when communicating matrices and vectors to server processes (default where supported)
516: Level: developer
518: Note:
519: This is normally started automatically in `PetscInitialize()` when the option is provided
521: See `PCMPI` for information on using the solver with a `KSP` object
523: See `PetscShmgetAllocateArray()` for instructions on how to ensure the shared memory is available on your machine.
525: Developer Notes:
526: When called on MPI rank 0 this sets `PETSC_COMM_WORLD` to `PETSC_COMM_SELF` to allow a main program
527: written with `PETSC_COMM_WORLD` to run correctly on the single rank while all the ranks
528: (that would normally be sharing `PETSC_COMM_WORLD`) to run the solver server.
530: Can this be integrated into the `PetscDevice` abstraction that is currently being developed?
532: Conceivably `PCREDISTRIBUTE` could be organized in a similar manner to simplify its usage
534: This could be implemented directly at the `KSP` level instead of using the `PCMPI` wrapper object
536: The code could be extended to allow an MPI + OpenMP application to use the linear solver server concept across all shared-memory
537: nodes with a single MPI process per node for the user application but multiple MPI processes per node for the linear solver.
539: The concept could also be extended for users's callbacks for `SNES`, `TS`, and `Tao` where the `SNESSolve()` for example, runs on
540: all MPI processes but the user callback only runs on one MPI process per node.
542: PETSc could also be extended with an MPI-less API that provides access to PETSc's solvers without any reference to MPI, essentially remove
543: the `MPI_Comm` argument from PETSc calls.
545: .seealso: [](sec_pcmpi), `PCMPIServerEnd()`, `PCMPI`, `KSPCheckPCMPI()`
546: @*/
547: PetscErrorCode PCMPIServerBegin(void)
548: {
549: PetscMPIInt rank;
551: PetscFunctionBegin;
552: PetscCall(PetscInfo(NULL, "Starting MPI Linear Solver Server\n"));
553: if (PetscDefined(USE_SINGLE_LIBRARY)) {
554: PetscCall(VecInitializePackage());
555: PetscCall(MatInitializePackage());
556: PetscCall(DMInitializePackage());
557: PetscCall(PCInitializePackage());
558: PetscCall(KSPInitializePackage());
559: PetscCall(SNESInitializePackage());
560: PetscCall(TSInitializePackage());
561: PetscCall(TaoInitializePackage());
562: }
563: PetscCall(PetscLogStageRegister("PCMPI", &PCMPIStage));
564: PetscCall(PetscLogEventRegister("ServerDist", PC_CLASSID, &EventServerDist));
565: PetscCall(PetscLogEventRegister("ServerDistMPI", PC_CLASSID, &EventServerDistMPI));
567: if (!PetscDefined(HAVE_SHMGET)) PCMPIServerUseShmget = PETSC_FALSE;
568: PetscCall(PetscOptionsGetBool(NULL, NULL, "-mpi_linear_solver_server_use_shared_memory", &PCMPIServerUseShmget, NULL));
570: PetscCallMPI(MPI_Comm_rank(PC_MPI_COMM_WORLD, &rank));
571: if (PCMPIServerUseShmget) {
572: #if defined(PETSC_HAVE_PTHREAD_MUTEX)
573: PetscMPIInt size;
575: PetscCallMPI(MPI_Comm_size(PETSC_COMM_WORLD, &size));
576: if (size > 1) {
577: pthread_mutex_t *locks;
579: if (rank == 0) {
580: PCMPIServerActive = PETSC_TRUE;
581: PetscCall(PetscShmgetAllocateArray(size, sizeof(pthread_mutex_t), (void **)&locks));
582: }
583: PetscCall(PetscShmgetMapAddresses(PETSC_COMM_WORLD, 1, (const void **)&locks, (void **)&PCMPIServerLocks));
584: if (rank == 0) {
585: pthread_mutexattr_t attr;
587: pthread_mutexattr_init(&attr);
588: pthread_mutexattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
590: for (int i = 1; i < size; i++) {
591: pthread_mutex_init(&PCMPIServerLocks[i], &attr);
592: pthread_mutex_lock(&PCMPIServerLocks[i]);
593: }
594: }
595: PetscCallMPI(MPI_Barrier(PETSC_COMM_WORLD));
596: }
597: #endif
598: }
599: if (rank == 0) {
600: PETSC_COMM_WORLD = PETSC_COMM_SELF;
601: PCMPIServerActive = PETSC_TRUE;
602: PetscFunctionReturn(PETSC_SUCCESS);
603: }
605: while (PETSC_TRUE) {
606: PCMPICommand request = PCMPI_CREATE;
607: #if defined(PETSC_HAVE_PTHREAD_MUTEX)
608: PetscMPIInt dummy1 = 1, dummy2;
609: #endif
611: // TODO: can we broadcast the number of active ranks here so only the correct subset of processes waits on the later scatters?
612: #if defined(PETSC_HAVE_PTHREAD_MUTEX)
613: if (PCMPIServerUseShmget) pthread_mutex_lock(&PCMPIServerLocks[PetscGlobalRank]);
614: #endif
615: PetscCallMPI(MPI_Bcast(&request, 1, MPIU_ENUM, 0, PC_MPI_COMM_WORLD));
616: #if defined(PETSC_HAVE_PTHREAD_MUTEX)
617: if (PCMPIServerUseShmget) {
618: /* next line ensures PetscGlobalRank has locked before rank 0 can take the lock back */
619: PetscCallMPI(MPI_Reduce(&dummy1, &dummy2, 1, MPI_INT, MPI_SUM, 0, PC_MPI_COMM_WORLD));
620: pthread_mutex_unlock(&PCMPIServerLocks[PetscGlobalRank]);
621: }
622: #endif
623: switch (request) {
624: case PCMPI_CREATE:
625: PetscCall(PCMPICreate(NULL));
626: break;
627: case PCMPI_SET_MAT:
628: PetscCall(PCMPISetMat(NULL));
629: break;
630: case PCMPI_UPDATE_MAT_VALUES:
631: PetscCall(PCMPIUpdateMatValues(NULL));
632: break;
633: case PCMPI_VIEW:
634: // PetscCall(PCMPIView(NULL));
635: break;
636: case PCMPI_SOLVE:
637: PetscCall(PCMPISolve(NULL, NULL, NULL));
638: break;
639: case PCMPI_DESTROY:
640: PetscCall(PCMPIDestroy(NULL));
641: break;
642: case PCMPI_EXIT:
643: if (PCMPIServerUseShmget) PetscCall(PetscShmgetUnmapAddresses(1, (void **)&PCMPIServerLocks));
644: PetscCall(PetscFinalize());
645: exit(0); /* not sure if this is a good idea, but cannot return because it will run users main program */
646: break;
647: default:
648: break;
649: }
650: }
651: PetscFunctionReturn(PETSC_SUCCESS);
652: }
654: /*@C
655: PCMPIServerEnd - ends a server that runs on the rank != 0 MPI processes waiting to process requests for
656: parallel KSP solves and management of parallel `KSP` objects.
658: Logically Collective on all MPI ranks except 0
660: Level: developer
662: Note:
663: This is normally called automatically in `PetscFinalize()`
665: .seealso: [](sec_pcmpi), `PCMPIServerBegin()`, `PCMPI`, `KSPCheckPCMPI()`
666: @*/
667: PetscErrorCode PCMPIServerEnd(void)
668: {
669: PetscFunctionBegin;
670: if (PetscGlobalRank == 0) {
671: PetscViewer viewer = NULL;
672: PetscViewerFormat format;
674: PetscCall(PetscShmgetAddressesFinalize());
675: PetscCall(PCMPIServerBroadcastRequest(PCMPI_EXIT));
676: if (PCMPIServerUseShmget) PetscCall(PetscShmgetUnmapAddresses(1, (void **)&PCMPIServerLocks));
677: PETSC_COMM_WORLD = MPI_COMM_WORLD; /* could use PC_MPI_COMM_WORLD */
678: PetscOptionsBegin(PETSC_COMM_SELF, NULL, "MPI linear solver server options", NULL);
679: PetscCall(PetscOptionsViewer("-mpi_linear_solver_server_view", "View information about system solved with the server", "PCMPI", &viewer, &format, NULL));
680: PetscOptionsEnd();
681: if (viewer) {
682: PetscBool isascii;
684: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
685: if (isascii) {
686: PetscMPIInt size;
687: PetscMPIInt i;
689: PetscCallMPI(MPI_Comm_size(PETSC_COMM_WORLD, &size));
690: PetscCall(PetscViewerASCIIPrintf(viewer, "MPI linear solver server statistics:\n"));
691: PetscCall(PetscViewerASCIIPrintf(viewer, " Ranks KSPSolve()s Mats KSPs Avg. Size Avg. Its\n"));
692: if (PCMPIKSPCountsSeq) {
693: PetscCall(PetscViewerASCIIPrintf(viewer, " Sequential %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT "\n", PCMPISolveCountsSeq, PCMPIKSPCountsSeq, PCMPISizesSeq / PCMPISolveCountsSeq, PCMPIIterationsSeq / PCMPISolveCountsSeq));
694: }
695: for (i = 0; i < size; i++) {
696: if (PCMPIKSPCounts[i]) {
697: 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]));
698: }
699: }
700: PetscCall(PetscViewerASCIIPrintf(viewer, "MPI linear solver server %susing shared memory\n", PCMPIServerUseShmget ? "" : "not "));
701: }
702: PetscCall(PetscViewerDestroy(&viewer));
703: }
704: }
705: PetscCall(PCMPICommsDestroy());
706: PCMPIServerActive = PETSC_FALSE;
707: PetscFunctionReturn(PETSC_SUCCESS);
708: }
710: /*
711: This version is used in the trivial case when the MPI parallel solver server is running on just the original MPI rank 0
712: because, for example, the problem is small. This version is more efficient because it does not require copying any data
713: */
714: static PetscErrorCode PCSetUp_Seq(PC pc)
715: {
716: PC_MPI *km = (PC_MPI *)pc->data;
717: Mat sA;
718: const char *prefix;
719: char *found = NULL, *cprefix;
721: PetscFunctionBegin;
722: PCMPIServerInSolve = PETSC_TRUE;
723: PetscCall(PCGetOperators(pc, NULL, &sA));
724: PetscCall(PCGetOptionsPrefix(pc, &prefix));
725: PetscCall(KSPCreate(PETSC_COMM_SELF, &km->ksps[0]));
726: PetscCall(KSPSetNestLevel(km->ksps[0], 1));
727: PetscCall(PetscObjectSetTabLevel((PetscObject)km->ksps[0], 1));
729: /* Created KSP gets prefix of PC minus the mpi_linear_solver_server_ portion */
730: PetscCall(PCGetOptionsPrefix(pc, &prefix));
731: PetscCheck(prefix, PETSC_COMM_SELF, PETSC_ERR_PLIB, "PCMPI missing required prefix");
732: PetscCall(PetscStrallocpy(prefix, &cprefix));
733: PetscCall(PetscStrstr(cprefix, "mpi_linear_solver_server_", &found));
734: PetscCheck(found, PETSC_COMM_SELF, PETSC_ERR_PLIB, "PCMPI missing mpi_linear_solver_server_ portion of prefix");
735: *found = 0;
736: PetscCall(KSPSetOptionsPrefix(km->ksps[0], cprefix));
737: PetscCall(PetscFree(cprefix));
739: PetscCall(KSPSetOperators(km->ksps[0], sA, sA));
740: PetscCall(KSPSetFromOptions(km->ksps[0]));
741: PetscCall(KSPSetUp(km->ksps[0]));
742: PetscCall(PetscInfo(pc, "MPI parallel linear solver system is being solved directly on rank 0 due to its small size\n"));
743: PCMPIKSPCountsSeq++;
744: PCMPIServerInSolve = PETSC_FALSE;
745: PetscFunctionReturn(PETSC_SUCCESS);
746: }
748: static PetscErrorCode PCApply_Seq(PC pc, Vec b, Vec x)
749: {
750: PC_MPI *km = (PC_MPI *)pc->data;
751: PetscInt its, n;
752: Mat A;
754: PetscFunctionBegin;
755: PCMPIServerInSolve = PETSC_TRUE;
756: PetscCall(KSPSolve(km->ksps[0], b, x));
757: PetscCall(KSPGetIterationNumber(km->ksps[0], &its));
758: PCMPISolveCountsSeq++;
759: PCMPIIterationsSeq += its;
760: PetscCall(KSPGetOperators(km->ksps[0], NULL, &A));
761: PetscCall(MatGetSize(A, &n, NULL));
762: PCMPISizesSeq += n;
763: PCMPIServerInSolve = PETSC_FALSE;
764: /*
765: do not keep reference to previous rhs and solution since destroying them in the next KSPSolve()
766: my use PetscFree() instead of PCMPIArrayDeallocate()
767: */
768: PetscCall(VecDestroy(&km->ksps[0]->vec_rhs));
769: PetscCall(VecDestroy(&km->ksps[0]->vec_sol));
770: PetscFunctionReturn(PETSC_SUCCESS);
771: }
773: static PetscErrorCode PCView_Seq(PC pc, PetscViewer viewer)
774: {
775: PC_MPI *km = (PC_MPI *)pc->data;
777: PetscFunctionBegin;
778: PetscCall(PetscViewerASCIIPrintf(viewer, "Running MPI linear solver server directly on rank 0 due to its small size\n"));
779: PetscCall(PetscViewerASCIIPrintf(viewer, "Desired minimum number of nonzeros per rank for MPI parallel solve %" PetscInt_FMT "\n", km->mincntperrank));
780: PetscCall(PetscViewerASCIIPrintf(viewer, "*** Use -mpi_linear_solver_server_view to statistics on all the solves ***\n"));
781: PetscFunctionReturn(PETSC_SUCCESS);
782: }
784: static PetscErrorCode PCDestroy_Seq(PC pc)
785: {
786: PC_MPI *km = (PC_MPI *)pc->data;
787: Mat A, B;
788: Vec x, b;
790: PetscFunctionBegin;
791: PCMPIServerInSolve = PETSC_TRUE;
792: /* since matrices and vectors are shared with outer KSP we need to ensure they are not destroyed with PetscFree() */
793: PetscCall(KSPGetOperators(km->ksps[0], &A, &B));
794: PetscCall(PetscObjectReference((PetscObject)A));
795: PetscCall(PetscObjectReference((PetscObject)B));
796: PetscCall(KSPGetSolution(km->ksps[0], &x));
797: PetscCall(PetscObjectReference((PetscObject)x));
798: PetscCall(KSPGetRhs(km->ksps[0], &b));
799: PetscCall(PetscObjectReference((PetscObject)b));
800: PetscCall(KSPDestroy(&km->ksps[0]));
801: PetscCall(PetscFree(pc->data));
802: PCMPIServerInSolve = PETSC_FALSE;
803: PetscCall(MatDestroy(&A));
804: PetscCall(MatDestroy(&B));
805: PetscCall(VecDestroy(&x));
806: PetscCall(VecDestroy(&b));
807: PetscFunctionReturn(PETSC_SUCCESS);
808: }
810: /*
811: PCSetUp_MPI - Trigger the creation of the MPI parallel PC and copy parts of the matrix and
812: right-hand side to the parallel PC
813: */
814: static PetscErrorCode PCSetUp_MPI(PC pc)
815: {
816: PC_MPI *km = (PC_MPI *)pc->data;
817: PetscMPIInt rank, size;
818: PetscBool newmatrix = PETSC_FALSE;
820: PetscFunctionBegin;
821: PetscCallMPI(MPI_Comm_rank(MPI_COMM_WORLD, &rank));
822: 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?");
823: PetscCallMPI(MPI_Comm_size(MPI_COMM_WORLD, &size));
825: if (!pc->setupcalled) {
826: if (!km->alwaysuseserver) {
827: PetscInt n;
828: Mat sA;
829: /* short circuit for small systems */
830: PetscCall(PCGetOperators(pc, &sA, &sA));
831: PetscCall(MatGetSize(sA, &n, NULL));
832: if (n < 2 * km->mincntperrank - 1 || size == 1) {
833: pc->ops->setup = NULL;
834: pc->ops->apply = PCApply_Seq;
835: pc->ops->destroy = PCDestroy_Seq;
836: pc->ops->view = PCView_Seq;
837: PetscCall(PCSetUp_Seq(pc));
838: PetscFunctionReturn(PETSC_SUCCESS);
839: }
840: }
842: PetscCall(PCMPIServerBroadcastRequest(PCMPI_CREATE));
843: PetscCall(PCMPICreate(pc));
844: newmatrix = PETSC_TRUE;
845: }
846: if (pc->flag == DIFFERENT_NONZERO_PATTERN) newmatrix = PETSC_TRUE;
848: if (newmatrix) {
849: PetscCall(PetscInfo(pc, "New matrix or matrix has changed nonzero structure\n"));
850: PetscCall(PCMPIServerBroadcastRequest(PCMPI_SET_MAT));
851: PetscCall(PCMPISetMat(pc));
852: } else {
853: PetscCall(PetscInfo(pc, "Matrix has only changed nonzero values\n"));
854: PetscCall(PCMPIServerBroadcastRequest(PCMPI_UPDATE_MAT_VALUES));
855: PetscCall(PCMPIUpdateMatValues(pc));
856: }
857: PetscFunctionReturn(PETSC_SUCCESS);
858: }
860: static PetscErrorCode PCApply_MPI(PC pc, Vec b, Vec x)
861: {
862: PetscFunctionBegin;
863: PetscCall(PCMPIServerBroadcastRequest(PCMPI_SOLVE));
864: PetscCall(PCMPISolve(pc, b, x));
865: PetscFunctionReturn(PETSC_SUCCESS);
866: }
868: static PetscErrorCode PCDestroy_MPI(PC pc)
869: {
870: PetscFunctionBegin;
871: PetscCall(PCMPIServerBroadcastRequest(PCMPI_DESTROY));
872: PetscCall(PCMPIDestroy(pc));
873: PetscCall(PetscFree(pc->data));
874: PetscFunctionReturn(PETSC_SUCCESS);
875: }
877: /*
878: PCView_MPI - Cannot call view on the MPI parallel KSP because other ranks do not have access to the viewer, use options database
879: */
880: static PetscErrorCode PCView_MPI(PC pc, PetscViewer viewer)
881: {
882: PC_MPI *km = (PC_MPI *)pc->data;
883: MPI_Comm comm;
884: PetscMPIInt size;
886: PetscFunctionBegin;
887: PetscCall(PetscObjectGetComm((PetscObject)km->ksps[0], &comm));
888: PetscCallMPI(MPI_Comm_size(comm, &size));
889: PetscCall(PetscViewerASCIIPrintf(viewer, "Size of MPI communicator used for MPI parallel KSP solve %d\n", size));
890: PetscCall(PetscViewerASCIIPrintf(viewer, "Desired minimum number of matrix rows on each MPI process for MPI parallel solve %" PetscInt_FMT "\n", km->mincntperrank));
891: PetscCall(PetscViewerASCIIPrintf(viewer, "*** Use -mpi_linear_solver_server_view to view statistics on all the solves ***\n"));
892: PetscFunctionReturn(PETSC_SUCCESS);
893: }
895: static PetscErrorCode PCSetFromOptions_MPI(PC pc, PetscOptionItems PetscOptionsObject)
896: {
897: PC_MPI *km = (PC_MPI *)pc->data;
899: PetscFunctionBegin;
900: PetscOptionsHeadBegin(PetscOptionsObject, "MPI linear solver server options");
901: PetscCall(PetscOptionsInt("-minimum_count_per_rank", "Desired minimum number of nonzeros per rank", "None", km->mincntperrank, &km->mincntperrank, NULL));
902: 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));
903: PetscOptionsHeadEnd();
904: PetscFunctionReturn(PETSC_SUCCESS);
905: }
907: /*MC
908: PCMPI - Calls an MPI parallel `KSP` to solve a linear system from user code running on one process
910: Options Database Keys for the Server:
911: + -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
912: . -mpi_linear_solver_server_view - displays information about all the linear systems solved by the MPI linear solver server
913: - -mpi_linear_solver_server_use_shared_memory <true, false> - use shared memory to distribute matrix and right hand side, defaults to true
915: Options Database Keys for a specific `KSP` object
916: + -[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
917: - -[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)
919: Level: developer
921: Notes:
922: This cannot be used with vectors or matrices that are created using arrays provided by the user, such as `VecCreateWithArray()` or
923: `MatCreateSeqAIJWithArrays()`
925: 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.
927: It can be particularly useful for user OpenMP code or potentially user GPU code.
929: When the program is running with a single MPI process then it directly uses the provided matrix and right-hand side
930: 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.
932: 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
933: because they are not the actual solver objects.
935: 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
936: 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.
938: Developer Note:
939: 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
940: a `KSP`. The outer most `KSP` object is automatically set to `KSPPREONLY` and thus is not directly visible to the user.
942: .seealso: [](sec_pcmpi), `KSPCreate()`, `KSPSetType()`, `KSPType`, `KSP`, `PC`, `PCMPIServerBegin()`, `PCMPIServerEnd()`, `KSPCheckPCMPI()`
943: M*/
944: PETSC_EXTERN PetscErrorCode PCCreate_MPI(PC pc)
945: {
946: PC_MPI *km;
947: char *found = NULL;
949: PetscFunctionBegin;
950: PetscCall(PetscStrstr(((PetscObject)pc)->prefix, "mpi_linear_solver_server_", &found));
951: PetscCheck(found, PETSC_COMM_SELF, PETSC_ERR_PLIB, "PCMPI object prefix does not have mpi_linear_solver_server_");
953: /* material from PCSetType() */
954: PetscTryTypeMethod(pc, destroy);
955: pc->ops->destroy = NULL;
956: pc->data = NULL;
958: PetscCall(PetscFunctionListDestroy(&((PetscObject)pc)->qlist));
959: PetscCall(PetscMemzero(pc->ops, sizeof(struct _PCOps)));
960: pc->modifysubmatrices = NULL;
961: pc->modifysubmatricesP = NULL;
962: pc->setupcalled = 0;
964: PetscCall(PetscNew(&km));
965: pc->data = (void *)km;
967: km->mincntperrank = 10000;
969: pc->ops->setup = PCSetUp_MPI;
970: pc->ops->apply = PCApply_MPI;
971: pc->ops->destroy = PCDestroy_MPI;
972: pc->ops->view = PCView_MPI;
973: pc->ops->setfromoptions = PCSetFromOptions_MPI;
974: PetscCall(PetscObjectChangeTypeName((PetscObject)pc, PCMPI));
975: PetscFunctionReturn(PETSC_SUCCESS);
976: }
978: /*@
979: PCMPIGetKSP - Gets the `KSP` created by the `PCMPI`
981: Not Collective
983: Input Parameter:
984: . pc - the preconditioner context
986: Output Parameter:
987: . innerksp - the inner `KSP`
989: Level: advanced
991: .seealso: [](ch_ksp), `KSP`, `PCMPI`, `PCREDISTRIBUTE`
992: @*/
993: PetscErrorCode PCMPIGetKSP(PC pc, KSP *innerksp)
994: {
995: PC_MPI *red = (PC_MPI *)pc->data;
997: PetscFunctionBegin;
999: PetscAssertPointer(innerksp, 2);
1000: *innerksp = red->ksps[0];
1001: PetscFunctionReturn(PETSC_SUCCESS);
1002: }