From 65a820ae8981a84fe7fbac87c48482e9f82b35f4 Mon Sep 17 00:00:00 2001 From: Jordan Henderson Date: Fri, 25 Jan 2019 15:32:51 -0600 Subject: Fix some collective metadata read issues --- release_docs/RELEASE.txt | 16 +++ src/H5C.c | 55 ++------ src/H5CX.c | 4 +- src/H5Cimage.c | 4 +- src/H5Dchunk.c | 22 +--- src/H5Dmpio.c | 6 +- src/H5Z.c | 2 +- testpar/t_cache.c | 16 +-- testpar/t_coll_md_read.c | 336 +++++++++++++++++++++++++++++++++++++++++++++-- testpar/testphdf5.c | 5 +- testpar/testphdf5.h | 2 + 11 files changed, 381 insertions(+), 87 deletions(-) diff --git a/release_docs/RELEASE.txt b/release_docs/RELEASE.txt index 2dfcfc1..9e63045 100644 --- a/release_docs/RELEASE.txt +++ b/release_docs/RELEASE.txt @@ -264,6 +264,22 @@ Bug Fixes since HDF5-1.10.3 release Library ------- + - Fix hangs with collective metadata reads during chunked dataset I/O + + In the parallel library, it was discovered that when a particular + sequence of operations following a pattern of: + + "write to chunked dataset" -> "flush file" -> "read from dataset" + + occurred with collective metadata reads enabled, hangs could be + observed due to certain MPI ranks not participating in the collective + metadata reads. + + To fix the issue, collective metadata reads are now disabled during + chunked dataset raw data I/O. + + (JTH - 2019/02/11, HDFFV-10563, HDFFV-10688) + - Deleting attributes in dense storage The library aborts with "infinite loop closing library" after diff --git a/src/H5C.c b/src/H5C.c index b5dc6a5..70eb7ba 100644 --- a/src/H5C.c +++ b/src/H5C.c @@ -1482,31 +1482,17 @@ H5C_insert_entry(H5F_t * f, H5C__UPDATE_STATS_FOR_INSERTION(cache_ptr, entry_ptr) #ifdef H5_HAVE_PARALLEL - if(H5F_HAS_FEATURE(f, H5FD_FEAT_HAS_MPI)) { - coll_access = (H5P_USER_TRUE == f->coll_md_read ? TRUE : FALSE); - - /* If not explicitly disabled, get the cmdr setting from the API context */ - if(!coll_access && H5P_FORCE_FALSE != f->coll_md_read) - coll_access = H5CX_get_coll_metadata_read(); - } /* end if */ + if(H5F_HAS_FEATURE(f, H5FD_FEAT_HAS_MPI)) + coll_access = H5CX_get_coll_metadata_read(); entry_ptr->coll_access = coll_access; if(coll_access) { H5C__INSERT_IN_COLL_LIST(cache_ptr, entry_ptr, FAIL) /* Make sure the size of the collective entries in the cache remain in check */ - if(H5P_USER_TRUE == f->coll_md_read) { - if(cache_ptr->max_cache_size * 80 < cache_ptr->coll_list_size * 100) { - if(H5C_clear_coll_entries(cache_ptr, TRUE) < 0) - HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "can't clear collective metadata entries") - } /* end if */ - } /* end if */ - else { - if(cache_ptr->max_cache_size * 40 < cache_ptr->coll_list_size * 100) { - if(H5C_clear_coll_entries(cache_ptr, TRUE) < 0) - HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "can't clear collective metadata entries") - } /* end if */ - } /* end else */ + if(cache_ptr->max_cache_size * 80 < cache_ptr->coll_list_size * 100) + if(H5C_clear_coll_entries(cache_ptr, TRUE) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "can't clear collective metadata entries") } /* end if */ #endif @@ -2242,13 +2228,8 @@ H5C_protect(H5F_t * f, ring = H5CX_get_ring(); #ifdef H5_HAVE_PARALLEL - if(H5F_HAS_FEATURE(f, H5FD_FEAT_HAS_MPI)) { - coll_access = (H5P_USER_TRUE == f->coll_md_read ? TRUE : FALSE); - - /* If not explicitly disabled, get the cmdr setting from the API context */ - if(!coll_access && H5P_FORCE_FALSE != f->coll_md_read) - coll_access = H5CX_get_coll_metadata_read(); - } /* end if */ + if(H5F_HAS_FEATURE(f, H5FD_FEAT_HAS_MPI)) + coll_access = H5CX_get_coll_metadata_read(); #endif /* H5_HAVE_PARALLEL */ /* first check to see if the target is in cache */ @@ -2285,7 +2266,7 @@ H5C_protect(H5F_t * f, the entry in their cache still have to participate in the bcast. */ #ifdef H5_HAVE_PARALLEL - if(H5F_HAS_FEATURE(f, H5FD_FEAT_HAS_MPI) && coll_access) { + if(coll_access) { if(!(entry_ptr->is_dirty) && !(entry_ptr->coll_access)) { MPI_Comm comm; /* File MPI Communicator */ int mpi_code; /* MPI error code */ @@ -2600,21 +2581,11 @@ H5C_protect(H5F_t * f, } /* end if */ #ifdef H5_HAVE_PARALLEL - if(H5F_HAS_FEATURE(f, H5FD_FEAT_HAS_MPI)) { - /* Make sure the size of the collective entries in the cache remain in check */ - if(coll_access) { - if(H5P_USER_TRUE == f->coll_md_read) { - if(cache_ptr->max_cache_size * 80 < cache_ptr->coll_list_size * 100) - if(H5C_clear_coll_entries(cache_ptr, TRUE) < 0) - HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, NULL, "can't clear collective metadata entries") - } /* end if */ - else { - if(cache_ptr->max_cache_size * 40 < cache_ptr->coll_list_size * 100) - if(H5C_clear_coll_entries(cache_ptr, TRUE) < 0) - HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, NULL, "can't clear collective metadata entries") - } /* end else */ - } /* end if */ - } /* end if */ + /* Make sure the size of the collective entries in the cache remain in check */ + if(coll_access) + if(cache_ptr->max_cache_size * 80 < cache_ptr->coll_list_size * 100) + if(H5C_clear_coll_entries(cache_ptr, TRUE) < 0) + HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, NULL, "can't clear collective metadata entries") #endif /* H5_HAVE_PARALLEL */ done: diff --git a/src/H5CX.c b/src/H5CX.c index 0d20132..f945df0 100644 --- a/src/H5CX.c +++ b/src/H5CX.c @@ -815,8 +815,8 @@ H5CX_set_apl(hid_t *acspl_id, const H5P_libclass_t *libclass, #ifdef H5_HAVE_PARALLEL /* If this routine is not guaranteed to be collective (i.e. it doesn't - * modify the structural metadata in a file), check if we should use - * a collective metadata read. + * modify the structural metadata in a file), check if the application + * specified a collective metadata read for just this operation. */ if(!is_collective) { H5P_genplist_t *plist; /* Property list pointer */ diff --git a/src/H5Cimage.c b/src/H5Cimage.c index bdab1a3..4684630 100644 --- a/src/H5Cimage.c +++ b/src/H5Cimage.c @@ -1065,9 +1065,9 @@ H5C__read_cache_image(H5F_t *f, H5C_t *cache_ptr) H5C__UPDATE_STATS_FOR_CACHE_IMAGE_READ(cache_ptr) #ifdef H5_HAVE_PARALLEL - if ( aux_ptr ) { + if ( aux_ptr ) { - /* Broadcast cache image */ + /* Broadcast cache image */ if ( MPI_SUCCESS != (mpi_result = MPI_Bcast(cache_ptr->image_buffer, (int)cache_ptr->image_len, MPI_BYTE, diff --git a/src/H5Dchunk.c b/src/H5Dchunk.c index e165569..af8ff91 100644 --- a/src/H5Dchunk.c +++ b/src/H5Dchunk.c @@ -2914,9 +2914,6 @@ H5D__chunk_lookup(const H5D_t *dset, const hsize_t *scaled, /* Check for cached information */ if(!H5D__chunk_cinfo_cache_found(&dset->shared->cache.chunk.last, udata)) { H5D_chk_idx_info_t idx_info; /* Chunked index info */ -#ifdef H5_HAVE_PARALLEL - H5P_coll_md_read_flag_t temp_cmr; /* Temp value to hold the coll metadata read setting */ -#endif /* H5_HAVE_PARALLEL */ /* Compose chunked index info struct */ idx_info.f = dset->oloc.file; @@ -2925,25 +2922,18 @@ H5D__chunk_lookup(const H5D_t *dset, const hsize_t *scaled, idx_info.storage = &dset->shared->layout.storage.u.chunk; #ifdef H5_HAVE_PARALLEL - if(H5F_HAS_FEATURE(idx_info.f, H5FD_FEAT_HAS_MPI)) { - /* disable collective metadata read for chunk indexes - as it is highly unlikely that users would read the - same chunks from all processes. MSC - might turn on - for root node? */ - temp_cmr = H5F_COLL_MD_READ(idx_info.f); - H5F_set_coll_md_read(idx_info.f, H5P_FORCE_FALSE); - } /* end if */ + /* Disable collective metadata read for chunk indexes as it is + * highly unlikely that users would read the same chunks from all + * processes. + */ + if(H5F_HAS_FEATURE(idx_info.f, H5FD_FEAT_HAS_MPI)) + H5CX_set_coll_metadata_read(FALSE); #endif /* H5_HAVE_PARALLEL */ /* Go get the chunk information */ if((dset->shared->layout.storage.u.chunk.ops->get_addr)(&idx_info, udata) < 0) HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't query chunk address") -#ifdef H5_HAVE_PARALLEL - if(H5F_HAS_FEATURE(idx_info.f, H5FD_FEAT_HAS_MPI)) - H5F_set_coll_md_read(idx_info.f, temp_cmr); -#endif /* H5_HAVE_PARALLEL */ - /* * Cache the information retrieved. * diff --git a/src/H5Dmpio.c b/src/H5Dmpio.c index f5da33d..0423006 100644 --- a/src/H5Dmpio.c +++ b/src/H5Dmpio.c @@ -800,6 +800,10 @@ H5D__chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_inf HDassert(type_info); HDassert(fm); + /* Disable collective metadata reads for chunked dataset I/O operations + * in order to prevent potential hangs */ + H5CX_set_coll_metadata_read(FALSE); + /* Check the optional property list for the collective chunk IO optimization option */ if(H5CX_get_mpio_chunk_opt_mode(&chunk_opt_mode) < 0) HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "couldn't get chunk optimization option") @@ -2313,7 +2317,7 @@ if(H5DEBUG(D)) /* Broadcasting the MPI_IO option info. and chunk address info. */ if(MPI_SUCCESS != (mpi_code = MPI_Bcast(total_chunk_addr_array, (int)(sizeof(haddr_t) * fm->layout->u.chunk.nchunks), MPI_BYTE, (int)0, io_info->comm))) - HMPI_GOTO_ERROR(FAIL, "MPI_BCast failed", mpi_code) + HMPI_GOTO_ERROR(FAIL, "MPI_BCast failed", mpi_code) } /* end if */ /* Start at first node in chunk skip list */ diff --git a/src/H5Z.c b/src/H5Z.c index f41588f..0e2a7ba 100644 --- a/src/H5Z.c +++ b/src/H5Z.c @@ -611,7 +611,7 @@ H5Z__flush_file_cb(void *obj_ptr, hid_t H5_ATTR_UNUSED obj_id, void *key) /* Sanity check for collectively calling H5Zunregister, if requested */ /* (Sanity check assumes that a barrier on one file's comm * is sufficient (i.e. that there aren't different comms for - * different files). -QAK, 2018/02/14 + * different files). -QAK, 2018/02/14) */ if(H5_coll_api_sanity_check_g && !object->sanity_checked) { MPI_Comm mpi_comm; /* File's communicator */ diff --git a/testpar/t_cache.c b/testpar/t_cache.c index 7488728..41c95e2 100644 --- a/testpar/t_cache.c +++ b/testpar/t_cache.c @@ -7225,7 +7225,7 @@ smoke_check_6(int metadata_write_strategy) /* some error occured in the server -- report failure */ nerrors++; if ( verbose ) { - HDfprintf(stdout, "%d:%s: server_main() failed.\n", + HDfprintf(stdout, "%d:%s: server_main() failed.\n", world_mpi_rank, FUNC); } } @@ -7241,7 +7241,7 @@ smoke_check_6(int metadata_write_strategy) fid = -1; cache_ptr = NULL; if ( verbose ) { - HDfprintf(stdout, "%d:%s: setup_cache_for_test() failed.\n", + HDfprintf(stdout, "%d:%s: setup_cache_for_test() failed.\n", world_mpi_rank, FUNC); } } @@ -7250,7 +7250,7 @@ smoke_check_6(int metadata_write_strategy) virt_num_data_entries = NUM_DATA_ENTRIES; /* insert the first half collectively */ - file_ptr->coll_md_read = H5P_USER_TRUE; + H5CX_set_coll_metadata_read(TRUE); for ( i = 0; i < virt_num_data_entries/2; i++ ) { struct datum * entry_ptr; @@ -7271,7 +7271,7 @@ smoke_check_6(int metadata_write_strategy) } /* insert the other half independently */ - file_ptr->coll_md_read = H5P_USER_FALSE; + H5CX_set_coll_metadata_read(FALSE); for ( i = virt_num_data_entries/2; i < virt_num_data_entries; i++ ) { struct datum * entry_ptr; @@ -7291,7 +7291,7 @@ smoke_check_6(int metadata_write_strategy) HDassert(cache_ptr->max_cache_size*0.8 > cache_ptr->coll_list_size); } - /* flush the file */ + /* flush the file */ if ( H5Fflush(fid, H5F_SCOPE_GLOBAL) < 0 ) { nerrors++; if ( verbose ) { @@ -7301,7 +7301,7 @@ smoke_check_6(int metadata_write_strategy) } /* Protect the first half of the entries collectively */ - file_ptr->coll_md_read = H5P_USER_TRUE; + H5CX_set_coll_metadata_read(TRUE); for ( i = 0; i < (virt_num_data_entries / 2); i++ ) { struct datum * entry_ptr; @@ -7322,13 +7322,13 @@ smoke_check_6(int metadata_write_strategy) } /* protect the other half independently */ - file_ptr->coll_md_read = H5P_USER_FALSE; + H5CX_set_coll_metadata_read(FALSE); for ( i = virt_num_data_entries/2; i < virt_num_data_entries; i++ ) { struct datum * entry_ptr; entry_ptr = &(data[i]); - lock_entry(file_ptr, i); + lock_entry(file_ptr, i); if(FALSE != entry_ptr->header.coll_access) { nerrors++; diff --git a/testpar/t_coll_md_read.c b/testpar/t_coll_md_read.c index f945d2b..912388c 100644 --- a/testpar/t_coll_md_read.c +++ b/testpar/t_coll_md_read.c @@ -32,6 +32,14 @@ #define PARTIAL_NO_SELECTION_Y_DIM_SCALE 5 #define PARTIAL_NO_SELECTION_X_DIM_SCALE 5 +#define MULTI_CHUNK_IO_ADDRMAP_ISSUE_DIMS 2 + +#define LINK_CHUNK_IO_SORT_CHUNK_ISSUE_NO_SEL_PROCESS (mpi_rank == mpi_size - 1) +#define LINK_CHUNK_IO_SORT_CHUNK_ISSUE_DATASET_NAME "linked_chunk_io_sort_chunk_issue" +#define LINK_CHUNK_IO_SORT_CHUNK_ISSUE_Y_DIM_SCALE 20000 +#define LINK_CHUNK_IO_SORT_CHUNK_ISSUE_CHUNK_SIZE 1 +#define LINK_CHUNK_IO_SORT_CHUNK_ISSUE_DIMS 1 + /* * A test for issue HDFFV-10501. A parallel hang was reported which occurred * in linked-chunk I/O when collective metadata reads are enabled and some ranks @@ -57,13 +65,13 @@ void test_partial_no_selection_coll_md_read(void) hsize_t stride[PARTIAL_NO_SELECTION_DATASET_NDIMS]; hsize_t count[PARTIAL_NO_SELECTION_DATASET_NDIMS]; hsize_t block[PARTIAL_NO_SELECTION_DATASET_NDIMS]; - hid_t file_id = -1; - hid_t fapl_id = -1; - hid_t dset_id = -1; - hid_t dcpl_id = -1; - hid_t dxpl_id = -1; - hid_t fspace_id = -1; - hid_t mspace_id = -1; + hid_t file_id = H5I_INVALID_HID; + hid_t fapl_id = H5I_INVALID_HID; + hid_t dset_id = H5I_INVALID_HID; + hid_t dcpl_id = H5I_INVALID_HID; + hid_t dxpl_id = H5I_INVALID_HID; + hid_t fspace_id = H5I_INVALID_HID; + hid_t mspace_id = H5I_INVALID_HID; int mpi_rank, mpi_size; void *data = NULL; void *read_buf = NULL; @@ -86,7 +94,7 @@ void test_partial_no_selection_coll_md_read(void) file_id = H5Fcreate(filename, H5F_ACC_TRUNC, H5P_DEFAULT, fapl_id); VRFY((file_id >= 0), "H5Fcreate succeeded"); - dataset_dims = malloc(PARTIAL_NO_SELECTION_DATASET_NDIMS * sizeof(*dataset_dims)); + dataset_dims = HDmalloc(PARTIAL_NO_SELECTION_DATASET_NDIMS * sizeof(*dataset_dims)); VRFY((dataset_dims != NULL), "malloc succeeded"); dataset_dims[0] = PARTIAL_NO_SELECTION_Y_DIM_SCALE * mpi_size; @@ -129,7 +137,7 @@ void test_partial_no_selection_coll_md_read(void) mspace_id = H5Screate_simple(1, sel_dims, NULL); VRFY((mspace_id >= 0), "H5Screate_simple succeeded"); - data = calloc(1, count[1] * (PARTIAL_NO_SELECTION_Y_DIM_SCALE * PARTIAL_NO_SELECTION_X_DIM_SCALE) * sizeof(int)); + data = HDcalloc(1, count[1] * (PARTIAL_NO_SELECTION_Y_DIM_SCALE * PARTIAL_NO_SELECTION_X_DIM_SCALE) * sizeof(int)); VRFY((data != NULL), "calloc succeeded"); dxpl_id = H5Pcreate(H5P_DATASET_XFER); @@ -151,7 +159,7 @@ void test_partial_no_selection_coll_md_read(void) */ VRFY((H5Pset_dxpl_mpio_chunk_opt(dxpl_id, H5FD_MPIO_CHUNK_ONE_IO) >= 0), "H5Pset_dxpl_mpio_chunk_opt succeeded"); - read_buf = malloc(count[1] * (PARTIAL_NO_SELECTION_Y_DIM_SCALE * PARTIAL_NO_SELECTION_X_DIM_SCALE) * sizeof(int)); + read_buf = HDmalloc(count[1] * (PARTIAL_NO_SELECTION_Y_DIM_SCALE * PARTIAL_NO_SELECTION_X_DIM_SCALE) * sizeof(int)); VRFY((read_buf != NULL), "malloc succeeded"); /* @@ -171,21 +179,321 @@ void test_partial_no_selection_coll_md_read(void) * Check data integrity just to be sure. */ if (!PARTIAL_NO_SELECTION_NO_SEL_PROCESS) { - VRFY((!memcmp(data, read_buf, count[1] * (PARTIAL_NO_SELECTION_Y_DIM_SCALE * PARTIAL_NO_SELECTION_X_DIM_SCALE) * sizeof(int))), "memcmp succeeded"); + VRFY((!HDmemcmp(data, read_buf, count[1] * (PARTIAL_NO_SELECTION_Y_DIM_SCALE * PARTIAL_NO_SELECTION_X_DIM_SCALE) * sizeof(int))), "memcmp succeeded"); + } + + if (dataset_dims) { + HDfree(dataset_dims); + dataset_dims = NULL; } + if (data) { + HDfree(data); + data = NULL; + } + + if (read_buf) { + HDfree(read_buf); + read_buf = NULL; + } + + VRFY((H5Sclose(fspace_id) >= 0), "H5Sclose succeeded"); + VRFY((H5Sclose(mspace_id) >= 0), "H5Sclose succeeded"); + VRFY((H5Pclose(dcpl_id) >= 0), "H5Pclose succeeded"); + VRFY((H5Pclose(dxpl_id) >= 0), "H5Pclose succeeded"); + VRFY((H5Dclose(dset_id) >= 0), "H5Dclose succeeded"); + VRFY((H5Pclose(fapl_id) >= 0), "H5Pclose succeeded"); + VRFY((H5Fclose(file_id) >= 0), "H5Fclose succeeded"); +} + +/* + * A test for HDFFV-10562 which attempts to verify that using multi-chunk + * I/O with collective metadata reads enabled doesn't causes issues due to + * collective metadata reads being made only by process 0 in H5D__chunk_addrmap(). + * + * Failure in this test may either cause a hang, or, due to how the MPI calls + * pertaining to this issue might mistakenly match up, may cause an MPI error + * message similar to: + * + * #008: H5Dmpio.c line 2546 in H5D__obtain_mpio_mode(): MPI_BCast failed + * major: Internal error (too specific to document in detail) + * minor: Some MPI function failed + * #009: H5Dmpio.c line 2546 in H5D__obtain_mpio_mode(): Message truncated, error stack: + *PMPI_Bcast(1600)..................: MPI_Bcast(buf=0x1df98e0, count=18, MPI_BYTE, root=0, comm=0x84000006) failed + *MPIR_Bcast_impl(1452).............: + *MPIR_Bcast(1476)..................: + *MPIR_Bcast_intra(1249)............: + *MPIR_SMP_Bcast(1088)..............: + *MPIR_Bcast_binomial(239)..........: + *MPIDI_CH3U_Receive_data_found(131): Message from rank 0 and tag 2 truncated; 2616 bytes received but buffer size is 18 + * major: Internal error (too specific to document in detail) + * minor: MPI Error String + * + */ +void test_multi_chunk_io_addrmap_issue(void) +{ + const char *filename; + hsize_t start[MULTI_CHUNK_IO_ADDRMAP_ISSUE_DIMS]; + hsize_t stride[MULTI_CHUNK_IO_ADDRMAP_ISSUE_DIMS]; + hsize_t count[MULTI_CHUNK_IO_ADDRMAP_ISSUE_DIMS]; + hsize_t block[MULTI_CHUNK_IO_ADDRMAP_ISSUE_DIMS]; + hsize_t dims[MULTI_CHUNK_IO_ADDRMAP_ISSUE_DIMS] = {10, 5}; + hsize_t chunk_dims[MULTI_CHUNK_IO_ADDRMAP_ISSUE_DIMS] = {5, 5}; + hsize_t max_dims[MULTI_CHUNK_IO_ADDRMAP_ISSUE_DIMS] = {H5S_UNLIMITED, H5S_UNLIMITED}; + hid_t file_id = H5I_INVALID_HID; + hid_t fapl_id = H5I_INVALID_HID; + hid_t dset_id = H5I_INVALID_HID; + hid_t dcpl_id = H5I_INVALID_HID; + hid_t dxpl_id = H5I_INVALID_HID; + hid_t space_id = H5I_INVALID_HID; + void *read_buf = NULL; + int mpi_rank; + int data[5][5] = { {0, 1, 2, 3, 4}, + {0, 1, 2, 3, 4}, + {0, 1, 2, 3, 4}, + {0, 1, 2, 3, 4}, + {0, 1, 2, 3, 4} }; + + MPI_Comm_rank(MPI_COMM_WORLD, &mpi_rank); + + filename = GetTestParameters(); + + fapl_id = create_faccess_plist(MPI_COMM_WORLD, MPI_INFO_NULL, facc_type); + VRFY((fapl_id >= 0), "create_faccess_plist succeeded"); + + /* + * Even though the testphdf5 framework currently sets collective metadata reads + * on the FAPL, we call it here just to be sure this is futureproof, since + * demonstrating this issue relies upon it. + */ + VRFY((H5Pset_all_coll_metadata_ops(fapl_id, true) >= 0), "Set collective metadata reads succeeded"); + + file_id = H5Fcreate(filename, H5F_ACC_TRUNC, H5P_DEFAULT, fapl_id); + VRFY((file_id >= 0), "H5Fcreate succeeded"); + + space_id = H5Screate_simple(MULTI_CHUNK_IO_ADDRMAP_ISSUE_DIMS, dims, max_dims); + VRFY((space_id >= 0), "H5Screate_simple succeeded"); + + dcpl_id = H5Pcreate(H5P_DATASET_CREATE); + VRFY((dcpl_id >= 0), "H5Pcreate succeeded"); + + VRFY((H5Pset_chunk(dcpl_id, MULTI_CHUNK_IO_ADDRMAP_ISSUE_DIMS, chunk_dims) >= 0), "H5Pset_chunk succeeded"); + + dset_id = H5Dcreate2(file_id, "dset", H5T_NATIVE_INT, space_id, H5P_DEFAULT, dcpl_id, H5P_DEFAULT); + VRFY((dset_id >= 0), "H5Dcreate2 succeeded"); + + dxpl_id = H5Pcreate(H5P_DATASET_XFER); + VRFY((dxpl_id >= 0), "H5Pcreate succeeded"); + + VRFY((H5Pset_dxpl_mpio(dxpl_id, H5FD_MPIO_COLLECTIVE) >= 0), "H5Pset_dxpl_mpio succeeded"); + VRFY((H5Pset_dxpl_mpio_chunk_opt(dxpl_id, H5FD_MPIO_CHUNK_MULTI_IO) >= 0), "H5Pset_dxpl_mpio_chunk_opt succeeded"); + + start[1] = 0; + stride[0] = stride[1] = 1; + count[0] = count[1] = 5; + block[0] = block[1] = 1; + + if (mpi_rank == 0) + start[0] = 0; + else + start[0] = 5; + + VRFY((H5Sselect_hyperslab(space_id, H5S_SELECT_SET, start, stride, count, block) >= 0), "H5Sselect_hyperslab succeeded"); + if (mpi_rank != 0) + VRFY((H5Sselect_none(space_id) >= 0), "H5Sselect_none succeeded"); + + VRFY((H5Dwrite(dset_id, H5T_NATIVE_INT, H5S_ALL, space_id, dxpl_id, data) >= 0), "H5Dwrite succeeded"); + + VRFY((H5Fflush(file_id, H5F_SCOPE_GLOBAL) >= 0), "H5Fflush succeeded"); + + read_buf = HDmalloc(50 * sizeof(int)); + VRFY((read_buf != NULL), "malloc succeeded"); + + VRFY((H5Dread(dset_id, H5T_NATIVE_INT, H5S_ALL, H5S_ALL, dxpl_id, read_buf) >= 0), "H5Dread succeeded"); + + if (read_buf) { + HDfree(read_buf); + read_buf = NULL; + } + + VRFY((H5Sclose(space_id) >= 0), "H5Sclose succeeded"); + VRFY((H5Pclose(dcpl_id) >= 0), "H5Pclose succeeded"); + VRFY((H5Pclose(dxpl_id) >= 0), "H5Pclose succeeded"); + VRFY((H5Dclose(dset_id) >= 0), "H5Dclose succeeded"); + VRFY((H5Pclose(fapl_id) >= 0), "H5Pclose succeeded"); + VRFY((H5Fclose(file_id) >= 0), "H5Fclose succeeded"); +} + +/* + * A test for HDFFV-10562 which attempts to verify that using linked-chunk + * I/O with collective metadata reads enabled doesn't cause issues due to + * collective metadata reads being made only by process 0 in H5D__sort_chunk(). + * + * NOTE: Due to the way that the threshold value which pertains to this test + * is currently calculated within HDF5, there are several conditions that this + * test must maintain. Refer to the function H5D__sort_chunk in H5Dmpio.c for + * a better idea of why. + * + * Condition 1: We need to make sure that the test always selects every single + * chunk in the dataset. It is fine if the selection is split up among multiple + * ranks, but their combined selection must cover the whole dataset. + * + * Condition 2: The number of chunks in the dataset divided by the number of MPI + * ranks must exceed or equal 10000. In other words, each MPI rank must be + * responsible for 10000 or more unique chunks. + * + * Condition 3: This test will currently only be reliably reproducable for 2 or 3 + * MPI ranks. The threshold value calculated reduces to a constant 100 / mpi_size, + * and is compared against a default value of 30%. + * + * Failure in this test may either cause a hang, or, due to how the MPI calls + * pertaining to this issue might mistakenly match up, may cause an MPI error + * message similar to: + * + * #008: H5Dmpio.c line 2338 in H5D__sort_chunk(): MPI_BCast failed + * major: Internal error (too specific to document in detail) + * minor: Some MPI function failed + * #009: H5Dmpio.c line 2338 in H5D__sort_chunk(): Other MPI error, error stack: + *PMPI_Bcast(1600)........: MPI_Bcast(buf=0x7eae610, count=320000, MPI_BYTE, root=0, comm=0x84000006) failed + *MPIR_Bcast_impl(1452)...: + *MPIR_Bcast(1476)........: + *MPIR_Bcast_intra(1249)..: + *MPIR_SMP_Bcast(1088)....: + *MPIR_Bcast_binomial(250): message sizes do not match across processes in the collective routine: Received 2096 but expected 320000 + * major: Internal error (too specific to document in detail) + * minor: MPI Error String + */ +void test_link_chunk_io_sort_chunk_issue(void) +{ + const char *filename; + hsize_t *dataset_dims = NULL; + hsize_t max_dataset_dims[LINK_CHUNK_IO_SORT_CHUNK_ISSUE_DIMS]; + hsize_t sel_dims[1]; + hsize_t chunk_dims[LINK_CHUNK_IO_SORT_CHUNK_ISSUE_DIMS] = { LINK_CHUNK_IO_SORT_CHUNK_ISSUE_DIMS }; + hsize_t start[LINK_CHUNK_IO_SORT_CHUNK_ISSUE_DIMS]; + hsize_t stride[LINK_CHUNK_IO_SORT_CHUNK_ISSUE_DIMS]; + hsize_t count[LINK_CHUNK_IO_SORT_CHUNK_ISSUE_DIMS]; + hsize_t block[LINK_CHUNK_IO_SORT_CHUNK_ISSUE_DIMS]; + hid_t file_id = H5I_INVALID_HID; + hid_t fapl_id = H5I_INVALID_HID; + hid_t dset_id = H5I_INVALID_HID; + hid_t dcpl_id = H5I_INVALID_HID; + hid_t dxpl_id = H5I_INVALID_HID; + hid_t fspace_id = H5I_INVALID_HID; + hid_t mspace_id = H5I_INVALID_HID; + int mpi_rank, mpi_size; + void *data = NULL; + void *read_buf = NULL; + + MPI_Comm_rank(MPI_COMM_WORLD, &mpi_rank); + MPI_Comm_size(MPI_COMM_WORLD, &mpi_size); + + filename = GetTestParameters(); + + fapl_id = create_faccess_plist(MPI_COMM_WORLD, MPI_INFO_NULL, facc_type); + VRFY((fapl_id >= 0), "create_faccess_plist succeeded"); + + /* + * Even though the testphdf5 framework currently sets collective metadata reads + * on the FAPL, we call it here just to be sure this is futureproof, since + * demonstrating this issue relies upon it. + */ + VRFY((H5Pset_all_coll_metadata_ops(fapl_id, true) >= 0), "Set collective metadata reads succeeded"); + + file_id = H5Fcreate(filename, H5F_ACC_TRUNC, H5P_DEFAULT, fapl_id); + VRFY((file_id >= 0), "H5Fcreate succeeded"); + + dataset_dims = HDmalloc(LINK_CHUNK_IO_SORT_CHUNK_ISSUE_DIMS * sizeof(*dataset_dims)); + VRFY((dataset_dims != NULL), "malloc succeeded"); + + dataset_dims[0] = LINK_CHUNK_IO_SORT_CHUNK_ISSUE_CHUNK_SIZE * mpi_size * LINK_CHUNK_IO_SORT_CHUNK_ISSUE_Y_DIM_SCALE; + max_dataset_dims[0] = H5S_UNLIMITED; + + fspace_id = H5Screate_simple(LINK_CHUNK_IO_SORT_CHUNK_ISSUE_DIMS, dataset_dims, max_dataset_dims); + VRFY((fspace_id >= 0), "H5Screate_simple succeeded"); + + /* + * Set up chunking on the dataset in order to reproduce the problem. + */ + dcpl_id = H5Pcreate(H5P_DATASET_CREATE); + VRFY((dcpl_id >= 0), "H5Pcreate succeeded"); + + VRFY((H5Pset_chunk(dcpl_id, LINK_CHUNK_IO_SORT_CHUNK_ISSUE_DIMS, chunk_dims) >= 0), "H5Pset_chunk succeeded"); + + dset_id = H5Dcreate2(file_id, LINK_CHUNK_IO_SORT_CHUNK_ISSUE_DATASET_NAME, H5T_NATIVE_INT, fspace_id, H5P_DEFAULT, dcpl_id, H5P_DEFAULT); + VRFY((dset_id >= 0), "H5Dcreate2 succeeded"); + + /* + * Setup hyperslab selection to split the dataset among the ranks. + * + * The ranks will write rows across the dataset. + */ + stride[0] = LINK_CHUNK_IO_SORT_CHUNK_ISSUE_CHUNK_SIZE; + count[0] = (dataset_dims[0] / LINK_CHUNK_IO_SORT_CHUNK_ISSUE_CHUNK_SIZE) / mpi_size; + start[0] = count[0] * mpi_rank; + block[0] = LINK_CHUNK_IO_SORT_CHUNK_ISSUE_CHUNK_SIZE; + + VRFY((H5Sselect_hyperslab(fspace_id, H5S_SELECT_SET, start, stride, count, block) >= 0), "H5Sselect_hyperslab succeeded"); + + sel_dims[0] = count[0] * (LINK_CHUNK_IO_SORT_CHUNK_ISSUE_CHUNK_SIZE); + + mspace_id = H5Screate_simple(1, sel_dims, NULL); + VRFY((mspace_id >= 0), "H5Screate_simple succeeded"); + + data = HDcalloc(1, count[0] * (LINK_CHUNK_IO_SORT_CHUNK_ISSUE_CHUNK_SIZE) * sizeof(int)); + VRFY((data != NULL), "calloc succeeded"); + + dxpl_id = H5Pcreate(H5P_DATASET_XFER); + VRFY((dxpl_id >= 0), "H5Pcreate succeeded"); + + /* + * Enable collective access for the data transfer. + */ + VRFY((H5Pset_dxpl_mpio(dxpl_id, H5FD_MPIO_COLLECTIVE) >= 0), "H5Pset_dxpl_mpio succeeded"); + + VRFY((H5Dwrite(dset_id, H5T_NATIVE_INT, mspace_id, fspace_id, dxpl_id, data) >= 0), "H5Dwrite succeeded"); + + VRFY((H5Fflush(file_id, H5F_SCOPE_GLOBAL) >= 0), "H5Fflush succeeded"); + + /* + * Ensure that linked-chunk I/O is performed since this is + * the particular code path where the issue lies and we don't + * want the library doing multi-chunk I/O behind our backs. + */ + VRFY((H5Pset_dxpl_mpio_chunk_opt(dxpl_id, H5FD_MPIO_CHUNK_ONE_IO) >= 0), "H5Pset_dxpl_mpio_chunk_opt succeeded"); + + read_buf = HDmalloc(count[0] * (LINK_CHUNK_IO_SORT_CHUNK_ISSUE_CHUNK_SIZE) * sizeof(int)); + VRFY((read_buf != NULL), "malloc succeeded"); + + VRFY((H5Sselect_hyperslab(fspace_id, H5S_SELECT_SET, start, stride, count, block) >= 0), "H5Sselect_hyperslab succeeded"); + + sel_dims[0] = count[0] * (LINK_CHUNK_IO_SORT_CHUNK_ISSUE_CHUNK_SIZE); + + VRFY((H5Sclose(mspace_id) >= 0), "H5Sclose succeeded"); + + mspace_id = H5Screate_simple(1, sel_dims, NULL); + VRFY((mspace_id >= 0), "H5Screate_simple succeeded"); + + read_buf = HDrealloc(read_buf, count[0] * (LINK_CHUNK_IO_SORT_CHUNK_ISSUE_CHUNK_SIZE) * sizeof(int)); + VRFY((read_buf != NULL), "realloc succeeded"); + + /* + * Finally have each rank read their section of data back from the dataset. + */ + VRFY((H5Dread(dset_id, H5T_NATIVE_INT, mspace_id, fspace_id, dxpl_id, read_buf) >= 0), "H5Dread succeeded"); + if (dataset_dims) { - free(dataset_dims); + HDfree(dataset_dims); dataset_dims = NULL; } if (data) { - free(data); + HDfree(data); data = NULL; } if (read_buf) { - free(read_buf); + HDfree(read_buf); read_buf = NULL; } diff --git a/testpar/testphdf5.c b/testpar/testphdf5.c index 69b66ae..999e17a 100644 --- a/testpar/testphdf5.c +++ b/testpar/testphdf5.c @@ -549,7 +549,10 @@ int main(int argc, char **argv) AddTest("noselcollmdread", test_partial_no_selection_coll_md_read, NULL, "Collective Metadata read with some ranks having no selection", PARATESTFILE); - + AddTest("MC coll MD read", test_multi_chunk_io_addrmap_issue, NULL, + "Collective MD read with multi chunk I/O (H5D__chunk_addrmap)", PARATESTFILE); + AddTest("LC coll MD read", test_link_chunk_io_sort_chunk_issue, NULL, + "Collective MD read with link chunk I/O (H5D__sort_chunk)", PARATESTFILE); /* Display testing information */ TestInfo(argv[0]); diff --git a/testpar/testphdf5.h b/testpar/testphdf5.h index 176574e..4409221 100644 --- a/testpar/testphdf5.h +++ b/testpar/testphdf5.h @@ -295,6 +295,8 @@ void compress_readAll(void); #endif /* H5_HAVE_FILTER_DEFLATE */ void test_dense_attr(void); void test_partial_no_selection_coll_md_read(void); +void test_multi_chunk_io_addrmap_issue(void); +void test_link_chunk_io_sort_chunk_issue(void); /* commonly used prototypes */ hid_t create_faccess_plist(MPI_Comm comm, MPI_Info info, int l_facc_type); -- cgit v0.12 From 8f447bd3564a88738b9a26b1b0ac3bba785efd39 Mon Sep 17 00:00:00 2001 From: Jordan Henderson Date: Mon, 11 Feb 2019 19:15:56 -0600 Subject: Add RELEASE.txt note for collective metadata read fixes --- release_docs/RELEASE.txt | 16 ++++++++++++++++ 1 file changed, 16 insertions(+) diff --git a/release_docs/RELEASE.txt b/release_docs/RELEASE.txt index ef3bda0..316798d 100644 --- a/release_docs/RELEASE.txt +++ b/release_docs/RELEASE.txt @@ -264,6 +264,22 @@ Bug Fixes since HDF5-1.10.3 release Library ------- + - Fix hangs with collective metadata reads during chunked dataset I/O + + In the parallel library, it was discovered that when a particular + sequence of operations following a pattern of: + + "write to chunked dataset" -> "flush file" -> "read from dataset" + + occurred with collective metadata reads enabled, hangs could be + observed due to certain MPI ranks not participating in the collective + metadata reads. + + To fix the issue, collective metadata reads are now disabled during + chunked dataset raw data I/O. + + (JTH - 2019/02/11, HDFFV-10563, HDFFV-10688) + - Performance issue when closing an object The slow down is due to the search of the "tag_list" to find -- cgit v0.12