1 /* Copyright (c) 2010-2019. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
7 #ifndef SMPI_FILE_HPP_INCLUDED
8 #define SMPI_FILE_HPP_INCLUDED
9 #include "simgrid/plugins/file_system.h"
10 #include "smpi_comm.hpp"
11 #include "smpi_coll.hpp"
12 #include "smpi_datatype.hpp"
13 #include "smpi_info.hpp"
16 XBT_LOG_EXTERNAL_CATEGORY(smpi_pmpi);
23 simgrid::s4u::File* file_;
25 MPI_Offset* shared_file_pointer_;
26 s4u::MutexPtr shared_mutex_;
30 File(MPI_Comm comm, const char *filename, int amode, MPI_Info info);
31 File(const File&) = delete;
32 File& operator=(const File&) = delete;
35 int get_position(MPI_Offset* offset);
36 int get_position_shared(MPI_Offset* offset);
40 int seek(MPI_Offset offset, int whence);
41 int seek_shared(MPI_Offset offset, int whence);
43 void set_info( MPI_Info info);
44 static int read(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status);
45 static int read_shared(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status);
46 static int read_ordered(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status);
47 static int write(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status);
48 static int write_shared(MPI_File fh, const void *buf, int count,MPI_Datatype datatype, MPI_Status *status);
49 static int write_ordered(MPI_File fh, const void *buf, int count,MPI_Datatype datatype, MPI_Status *status);
50 template <int (*T)(MPI_File, void *, int, MPI_Datatype, MPI_Status *)> int op_all(void *buf, int count,MPI_Datatype datatype, MPI_Status *status);
51 static int close(MPI_File *fh);
52 static int del(const char *filename, MPI_Info info);
55 /* Read_all, Write_all : loosely based on */
56 /* @article{Thakur:1996:ETM:245875.245879,*/
57 /* author = {Thakur, Rajeev and Choudhary, Alok},*/
58 /* title = {An Extended Two-phase Method for Accessing Sections of Out-of-core Arrays},*/
59 /* journal = {Sci. Program.},*/
60 /* issue_date = {Winter 1996},*/
61 /* pages = {301--317},*/
63 template <int (*T)(MPI_File, void *, int, MPI_Datatype, MPI_Status *)>
64 int File::op_all(void *buf, int count, MPI_Datatype datatype, MPI_Status *status){
65 //get min and max offsets from everyone.
66 int size = comm_->size();
67 int rank = comm_-> rank();
68 MPI_Offset min_offset = file_->tell();
69 MPI_Offset max_offset = (min_offset + count * datatype->size());//cheating, as we don't care about exact data location, we can skip extent
70 MPI_Offset* min_offsets = new MPI_Offset[size];
71 MPI_Offset* max_offsets = new MPI_Offset[size];
72 simgrid::smpi::Colls::allgather(&min_offset, 1, MPI_OFFSET, min_offsets, 1, MPI_OFFSET, comm_);
73 simgrid::smpi::Colls::allgather(&max_offset, 1, MPI_OFFSET, max_offsets, 1, MPI_OFFSET, comm_);
74 MPI_Offset min=min_offset;
75 MPI_Offset max=max_offset;
78 for(int i=0;i<size;i++){
79 if(min_offsets[i]!=max_offsets[i])
81 tot+=(max_offsets[i]-min_offsets[i]);
82 if(min_offsets[i]<min)
84 if(max_offsets[i]>max)
88 XBT_CDEBUG(smpi_pmpi, "my offsets to read : %lld:%lld, global min and max %lld:%lld", min_offset, max_offset, min, max);
95 MPI_Offset total = max-min;
96 if(total==tot && (datatype->flags() & DT_FLAG_CONTIGUOUS)){
99 //contiguous. Just have each proc perform its read
100 status->count=count * datatype->size();
101 return T(this,buf,count,datatype, status);
104 //Interleaved case : How much do I need to read, and whom to send it ?
105 MPI_Offset my_chunk_start=(max-min+1)/size*rank;
106 MPI_Offset my_chunk_end=((max-min+1)/size*(rank+1));
107 XBT_CDEBUG(smpi_pmpi, "my chunks to read : %lld:%lld", my_chunk_start, my_chunk_end);
108 int* send_sizes = new int[size];
109 int* recv_sizes = new int[size];
110 int* send_disps = new int[size];
111 int* recv_disps = new int[size];
113 for(int i=0;i<size;i++){
115 send_disps[i]=0;//cheat to avoid issues when send>recv as we use recv buffer
116 if((my_chunk_start>=min_offsets[i] && my_chunk_start < max_offsets[i])||
117 ((my_chunk_end<=max_offsets[i]) && my_chunk_end> min_offsets[i])){
118 send_sizes[i]=(std::min(max_offsets[i]-1, my_chunk_end-1)-std::max(min_offsets[i], my_chunk_start));
119 //store min and max offest to actually read
120 min_offset=std::min(min_offset, min_offsets[i]);
121 total_sent+=send_sizes[i];
122 XBT_CDEBUG(smpi_pmpi, "will have to send %d bytes to %d", send_sizes[i], i);
125 min_offset=std::max(min_offset, my_chunk_start);
127 //merge the ranges of every process
128 std::vector<std::pair<MPI_Offset, MPI_Offset>> ranges;
129 for(int i=0; i<size; ++i)
130 ranges.push_back(std::make_pair(min_offsets[i],max_offsets[i]));
131 std::sort(ranges.begin(), ranges.end());
132 std::vector<std::pair<MPI_Offset, MPI_Offset>> chunks;
133 chunks.push_back(ranges[0]);
135 unsigned int nchunks=0;
137 while(i < ranges.size()){
138 if(ranges[i].second>chunks[nchunks].second){
139 // else range included - ignore
140 if(ranges[i].first>chunks[nchunks].second){
142 chunks.push_back(ranges[i]);
146 chunks[nchunks].second=ranges[i].second;
151 //what do I need to read ?
152 MPI_Offset totreads=0;
153 for(i=0; i<chunks.size();i++){
154 if(chunks[i].second < my_chunk_start)
156 else if (chunks[i].first > my_chunk_end)
159 totreads += (std::min(chunks[i].second, my_chunk_end-1)-std::max(chunks[i].first, my_chunk_start));
161 XBT_CDEBUG(smpi_pmpi, "will have to access %lld from my chunk", totreads);
163 unsigned char* sendbuf = smpi_get_tmp_sendbuffer(total_sent);
166 seek(min_offset, MPI_SEEK_SET);
167 T(this,sendbuf,totreads/datatype->size(),datatype, status);
169 simgrid::smpi::Colls::alltoall(send_sizes, 1, MPI_INT, recv_sizes, 1, MPI_INT, comm_);
171 for(int i=0;i<size;i++){
172 recv_disps[i]=total_recv;
173 total_recv+=recv_sizes[i];
175 //Set buf value to avoid copying dumb data
176 simgrid::smpi::Colls::alltoallv(sendbuf, send_sizes, send_disps, MPI_BYTE,
177 buf, recv_sizes, recv_disps, MPI_BYTE, comm_);
178 status->count=count * datatype->size();
179 smpi_free_tmp_buffer(sendbuf);
184 delete[] min_offsets;
185 delete[] max_offsets;