- Created by Barbara Jones, last modified on Aug 09, 2019
H5P_SET_DXPL_MPIO_CHUNK_OPT
Sets a flag specifying linked-chunk I/O or multi-chunk I/O
Procedure:
H5P_SET_DXPL_MPIO_CHUNK_OPT ( dxpl_id, opt_mode )
Signature:
herr_t H5Pset_dxpl_mpio_chunk_opt
(hid_t dxpl_id,
H5FD_mpio_chunk_opt_t opt_mode)
Parameters:
hid_t dxpl_id | IN: Data transfer property list identifier |
H5FD_mpio_chunk_opt_t opt_mode | IN: Optimization flag specifying linked-chunk I/O or multi-chunk I/O |
Description:
H5P_SET_DXPL_MPIO_CHUNK_OPT specifies whether I/O is to be performed as linked-chunk I/O or as multi-chunk I/O. This function overrides the HDF5 library's internal algorithm for determining which mechanism to use.
When an application uses collective I/O with chunked storage, the HDF5 library normally uses an internal algorithm to determine whether that I/O activity should be conducted as one linked-chunk I/O or as multi-chunk I/O. H5P_SET_DXPL_MPIO_CHUNK_OPT is provided so that an application can override the library's algorithm in circumstances where the library might lack the information needed to make an optimal decision.
H5P_SET_DXPL_MPIO_CHUNK_OPT works by setting one of the following flags in the parameter opt_mode:
H5FD_MPIO_CHUNK_ONE_IO | Do one link chunked I/O |
H5FD_MPIO_CHUNK_MULTI_IO | Do multi-chunked I/O |
This function works by setting a corresponding property in the dataset transfer property list dxpl_id.
The library performs I/O in the specified manner unless it determines that the low-level MPI IO package does not support the requested behavior; in such cases, the HDF5 library will internally use independent I/O.
Use of this function is optional.
Returns:
Returns a non-negative value if successful. Otherwise returns a negative value.
Example:
/* set up the collective transfer property list */
xfer_plist = H5Pcreate(H5P_DATASET_XFER);
VRFY((xfer_plist >= 0), "");
status = H5Pset_dxpl_mpio(xfer_plist, H5FD_MPIO_COLLECTIVE);
VRFY((status>= 0),"MPIO collective transfer property succeeded");
if(dxfer_coll_type == DXFER_INDEPENDENT_IO) {
status = H5Pset_dxpl_mpio_collective_opt(xfer_plist, H5FD_MPIO_INDIVIDUAL_IO);
VRFY((status>= 0),"set independent IO collectively succeeded");
}
switch(api_option){
case API_LINK_HARD:
status = H5Pset_dxpl_mpio_chunk_opt(xfer_plist,H5FD_MPIO_CHUNK_ONE_IO);
VRFY((status>= 0),"collective chunk optimization succeeded");
break;
case API_MULTI_HARD:
status = H5Pset_dxpl_mpio_chunk_opt(xfer_plist,H5FD_MPIO_CHUNK_MULTI_IO);
VRFY((status>= 0),"collective chunk optimization succeeded ");
break;
case API_LINK_TRUE:
status = H5Pset_dxpl_mpio_chunk_opt_num(xfer_plist,2);
VRFY((status>= 0),"collective chunk optimization set chunk number succeeded");
break;
case API_LINK_FALSE:
status = H5Pset_dxpl_mpio_chunk_opt_num(xfer_plist,6);
VRFY((status>= 0),"collective chunk optimization set chunk number succeeded");
break;
case API_MULTI_COLL:
status = H5Pset_dxpl_mpio_chunk_opt_num(xfer_plist,8);/* make sure it is using multi-chunk IO */
VRFY((status>= 0),"collective chunk optimization set chunk number succeeded");
status = H5Pset_dxpl_mpio_chunk_opt_ratio(xfer_plist,50);
VRFY((status>= 0),"collective chunk optimization set chunk ratio succeeded");
break;
case API_MULTI_IND:
status = H5Pset_dxpl_mpio_chunk_opt_num(xfer_plist,8);/* make sure it is using multi-chunk IO */
VRFY((status>= 0),"collective chunk optimization set chunk number succeeded");
status = H5Pset_dxpl_mpio_chunk_opt_ratio(xfer_plist,100);
VRFY((status>= 0),"collective chunk optimization set chunk ratio succeeded");
break;
default:
;
}
History:
--- Last Modified: August 09, 2019 | 02:02 PM