1 /* Copyright (c) 2010-2019. The SimGrid Team.
\r
2 * All rights reserved. */
\r
4 /* This program is free software; you can redistribute it and/or modify it
\r
5 * under the terms of the license (GNU LGPL) which comes with this package. */
\r
7 #ifndef SMPI_FILE_HPP_INCLUDED
\r
8 #define SMPI_FILE_HPP_INCLUDED
\r
9 #include "simgrid/plugins/file_system.h"
\r
10 #include "smpi_comm.hpp"
\r
11 #include "smpi_coll.hpp"
\r
12 #include "smpi_datatype.hpp"
\r
13 #include "smpi_info.hpp"
\r
14 #include <algorithm>
\r
22 simgrid::s4u::File* file_;
\r
24 MPI_Offset* shared_file_pointer_;
\r
25 s4u::MutexPtr shared_mutex_;
\r
29 File(MPI_Comm comm, char *filename, int amode, MPI_Info info);
\r
32 int get_position(MPI_Offset* offset);
\r
33 int get_position_shared(MPI_Offset* offset);
\r
36 int seek(MPI_Offset offset, int whence);
\r
37 int seek_shared(MPI_Offset offset, int whence);
\r
39 void set_info( MPI_Info info);
\r
40 static int read(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status);
\r
41 static int read_shared(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status);
\r
42 static int read_ordered(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status);
\r
43 static int write(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status);
\r
44 static int write_shared(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status);
\r
45 static int write_ordered(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status);
\r
46 template <int (*T)(MPI_File, void *, int, MPI_Datatype, MPI_Status *)> int op_all(void *buf, int count,MPI_Datatype datatype, MPI_Status *status);
\r
47 static int close(MPI_File *fh);
\r
48 static int del(char *filename, MPI_Info info);
\r
51 /* Read_all, Write_all : loosely based on */
\r
52 /* @article{Thakur:1996:ETM:245875.245879,*/
\r
53 /* author = {Thakur, Rajeev and Choudhary, Alok},*/
\r
54 /* title = {An Extended Two-phase Method for Accessing Sections of Out-of-core Arrays},*/
\r
55 /* journal = {Sci. Program.},*/
\r
56 /* issue_date = {Winter 1996},*/
\r
57 /* pages = {301--317},*/
\r
59 template <int (*T)(MPI_File, void *, int, MPI_Datatype, MPI_Status *)>
\r
60 int File::op_all(void *buf, int count, MPI_Datatype datatype, MPI_Status *status){
\r
61 //get min and max offsets from everyone.
\r
62 int size = comm_->size();
\r
63 int rank = comm_-> rank();
\r
64 MPI_Offset min_offset = file_->tell();
\r
65 MPI_Offset max_offset = (min_offset + count * datatype->size());//cheating, as we don't care about exact data location, we can skip extent
\r
66 MPI_Offset* min_offsets = xbt_new(MPI_Offset, size);
\r
67 MPI_Offset* max_offsets = xbt_new(MPI_Offset, size);
\r
68 simgrid::smpi::Colls::allgather(&min_offset, 1, MPI_OFFSET, min_offsets, 1, MPI_OFFSET, comm_);
\r
69 simgrid::smpi::Colls::allgather(&max_offset, 1, MPI_OFFSET, max_offsets, 1, MPI_OFFSET, comm_);
\r
70 MPI_Offset min=min_offset;
\r
71 MPI_Offset max=max_offset;
\r
74 for(int i=0;i<size;i++){
\r
75 if(min_offsets[i]!=max_offsets[i])
\r
77 tot+=(max_offsets[i]-min_offsets[i]);
\r
78 if(min_offsets[i]<min)
\r
80 if(max_offsets[i]>max)
\r
84 XBT_DEBUG("my offsets to read : %lld:%lld, global min and max %lld:%lld", min_offset, max_offset, min, max);
\r
89 MPI_Offset total = max-min;
\r
90 if(total==tot && (datatype->flags() & DT_FLAG_CONTIGUOUS)){
\r
91 //contiguous. Just have each proc perform its read
\r
92 status->count=count * datatype->size();
\r
93 return T(this,buf,count,datatype, status);
\r
96 //Interleaved case : How much do I need to read, and whom to send it ?
\r
97 MPI_Offset my_chunk_start=(max-min+1)/size*rank;
\r
98 MPI_Offset my_chunk_end=((max-min+1)/size*(rank+1));
\r
99 XBT_DEBUG("my chunks to read : %lld:%lld", my_chunk_start, my_chunk_end);
\r
100 int* send_sizes = xbt_new0(int, size);
\r
101 int* recv_sizes = xbt_new(int, size);
\r
102 int* send_disps = xbt_new(int, size);
\r
103 int* recv_disps = xbt_new(int, size);
\r
105 for(int i=0;i<size;i++){
\r
106 if((my_chunk_start>=min_offsets[i] && my_chunk_start < max_offsets[i])||
\r
107 ((my_chunk_end<=max_offsets[i]) && my_chunk_end> min_offsets[i])){
\r
108 send_sizes[i]=(std::min(max_offsets[i]-1, my_chunk_end-1)-std::max(min_offsets[i], my_chunk_start));
\r
109 //store min and max offest to actually read
\r
110 min_offset=std::min(min_offset, min_offsets[i]);
\r
111 send_disps[i]=0;//send_sizes[i]; cheat to avoid issues when send>recv as we use recv buffer
\r
112 total_sent+=send_sizes[i];
\r
113 XBT_DEBUG("will have to send %d bytes to %d", send_sizes[i], i);
\r
116 min_offset=std::max(min_offset, my_chunk_start);
\r
118 //merge the ranges of every process
\r
119 std::vector<std::pair<MPI_Offset, MPI_Offset>> ranges;
\r
120 for(int i=0; i<size; ++i)
\r
121 ranges.push_back(std::make_pair(min_offsets[i],max_offsets[i]));
\r
122 std::sort(ranges.begin(), ranges.end());
\r
123 std::vector<std::pair<MPI_Offset, MPI_Offset>> chunks;
\r
124 chunks.push_back(ranges[0]);
\r
126 unsigned int nchunks=0;
\r
128 while(i < ranges.size()){
\r
129 if(ranges[i].second>chunks[nchunks].second){
\r
130 // else range included - ignore
\r
131 if(ranges[i].first>chunks[nchunks].second){
\r
132 //new disjoint range
\r
133 chunks.push_back(ranges[i]);
\r
137 chunks[nchunks].second=ranges[i].second;
\r
142 //what do I need to read ?
\r
143 MPI_Offset totreads=0;
\r
144 for(i=0; i<chunks.size();i++){
\r
145 if(chunks[i].second < my_chunk_start)
\r
147 else if (chunks[i].first > my_chunk_end)
\r
150 totreads += (std::min(chunks[i].second, my_chunk_end-1)-std::max(chunks[i].first, my_chunk_start));
\r
152 XBT_DEBUG("will have to access %lld from my chunk", totreads);
\r
154 char* sendbuf= static_cast<char *>(smpi_get_tmp_sendbuffer(totreads));
\r
157 seek(min_offset, MPI_SEEK_SET);
\r
158 T(this,sendbuf,totreads/datatype->size(),datatype, status);
\r
160 simgrid::smpi::Colls::alltoall(send_sizes, 1, MPI_INT, recv_sizes, 1, MPI_INT, comm_);
\r
162 for(int i=0;i<size;i++){
\r
163 recv_disps[i]=total_recv;
\r
164 total_recv+=recv_sizes[i];
\r
166 //Set buf value to avoid copying dumb data
\r
167 simgrid::smpi::Colls::alltoallv(sendbuf, send_sizes, send_disps, MPI_BYTE,
\r
168 buf, recv_sizes, recv_disps, MPI_BYTE, comm_);
\r
169 status->count=count * datatype->size();
\r
170 smpi_free_tmp_buffer(sendbuf);
\r
171 xbt_free(send_sizes);
\r
172 xbt_free(recv_sizes);
\r
173 xbt_free(send_disps);
\r
174 xbt_free(recv_disps);
\r
175 xbt_free(min_offsets);
\r
176 xbt_free(max_offsets);
\r
177 return MPI_SUCCESS;
\r