-/* Copyright (c) 2010-2019. The SimGrid Team.\r
- * All rights reserved. */\r
-\r
-/* This program is free software; you can redistribute it and/or modify it\r
- * under the terms of the license (GNU LGPL) which comes with this package. */\r
-\r
-#ifndef SMPI_FILE_HPP_INCLUDED\r
-#define SMPI_FILE_HPP_INCLUDED\r
-#include "simgrid/plugins/file_system.h"\r
-#include "smpi_comm.hpp"\r
-#include "smpi_coll.hpp"\r
-#include "smpi_datatype.hpp"\r
-#include "smpi_info.hpp"\r
-#include <algorithm>\r
-\r
-\r
-namespace simgrid{\r
-namespace smpi{\r
-class File{\r
- MPI_Comm comm_;\r
- int flags_;\r
- simgrid::s4u::File* file_;\r
- MPI_Info info_;\r
- MPI_Offset* shared_file_pointer_;\r
- s4u::MutexPtr shared_mutex_;\r
- MPI_Win win_;\r
- char* list_;\r
- public:\r
- File(MPI_Comm comm, char *filename, int amode, MPI_Info info);\r
- ~File();\r
- int size();\r
- int get_position(MPI_Offset* offset);\r
- int get_position_shared(MPI_Offset* offset);\r
- int flags();\r
- int sync();\r
- int seek(MPI_Offset offset, int whence);\r
- int seek_shared(MPI_Offset offset, int whence);\r
- MPI_Info info();\r
- void set_info( MPI_Info info);\r
- static int read(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status);\r
- static int read_shared(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status);\r
- static int read_ordered(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status);\r
- static int write(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status);\r
- static int write_shared(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status);\r
- static int write_ordered(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status);\r
- 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
- static int close(MPI_File *fh);\r
- static int del(char *filename, MPI_Info info);\r
-};\r
-\r
- /* Read_all, Write_all : loosely based on */\r
- /* @article{Thakur:1996:ETM:245875.245879,*/\r
- /* author = {Thakur, Rajeev and Choudhary, Alok},*/\r
- /* title = {An Extended Two-phase Method for Accessing Sections of Out-of-core Arrays},*/\r
- /* journal = {Sci. Program.},*/\r
- /* issue_date = {Winter 1996},*/\r
- /* pages = {301--317},*/\r
- /* }*/ \r
- template <int (*T)(MPI_File, void *, int, MPI_Datatype, MPI_Status *)>\r
- int File::op_all(void *buf, int count, MPI_Datatype datatype, MPI_Status *status){\r
- //get min and max offsets from everyone.\r
- int size = comm_->size();\r
- int rank = comm_-> rank();\r
- MPI_Offset min_offset = file_->tell();\r
- MPI_Offset max_offset = (min_offset + count * datatype->size());//cheating, as we don't care about exact data location, we can skip extent\r
- MPI_Offset* min_offsets = xbt_new(MPI_Offset, size);\r
- MPI_Offset* max_offsets = xbt_new(MPI_Offset, size);\r
- simgrid::smpi::Colls::allgather(&min_offset, 1, MPI_OFFSET, min_offsets, 1, MPI_OFFSET, comm_);\r
- simgrid::smpi::Colls::allgather(&max_offset, 1, MPI_OFFSET, max_offsets, 1, MPI_OFFSET, comm_);\r
- MPI_Offset min=min_offset;\r
- MPI_Offset max=max_offset;\r
- MPI_Offset tot= 0;\r
- int empty=1;\r
- for(int i=0;i<size;i++){\r
- if(min_offsets[i]!=max_offsets[i])\r
- empty=0;\r
- tot+=(max_offsets[i]-min_offsets[i]);\r
- if(min_offsets[i]<min)\r
- min=min_offsets[i];\r
- if(max_offsets[i]>max)\r
- max=max_offsets[i];\r
- }\r
- \r
- XBT_DEBUG("my offsets to read : %lld:%lld, global min and max %lld:%lld", min_offset, max_offset, min, max);\r
- if(empty==1){\r
- status->count=0;\r
- return MPI_SUCCESS;\r
- }\r
- MPI_Offset total = max-min;\r
- if(total==tot && (datatype->flags() & DT_FLAG_CONTIGUOUS)){\r
- //contiguous. Just have each proc perform its read\r
- status->count=count * datatype->size();\r
- return T(this,buf,count,datatype, status);\r
- }\r
-\r
- //Interleaved case : How much do I need to read, and whom to send it ?\r
- MPI_Offset my_chunk_start=(max-min+1)/size*rank;\r
- MPI_Offset my_chunk_end=((max-min+1)/size*(rank+1));\r
- XBT_DEBUG("my chunks to read : %lld:%lld", my_chunk_start, my_chunk_end);\r
- int* send_sizes = xbt_new0(int, size);\r
- int* recv_sizes = xbt_new(int, size);\r
- int* send_disps = xbt_new(int, size);\r
- int* recv_disps = xbt_new(int, size);\r
- int total_sent=0;\r
- for(int i=0;i<size;i++){\r
- if((my_chunk_start>=min_offsets[i] && my_chunk_start < max_offsets[i])||\r
- ((my_chunk_end<=max_offsets[i]) && my_chunk_end> min_offsets[i])){\r
- send_sizes[i]=(std::min(max_offsets[i]-1, my_chunk_end-1)-std::max(min_offsets[i], my_chunk_start));\r
- //store min and max offest to actually read\r
- min_offset=std::min(min_offset, min_offsets[i]);\r
- send_disps[i]=0;//send_sizes[i]; cheat to avoid issues when send>recv as we use recv buffer\r
- total_sent+=send_sizes[i];\r
- XBT_DEBUG("will have to send %d bytes to %d", send_sizes[i], i);\r
- }\r
- }\r
- min_offset=std::max(min_offset, my_chunk_start);\r
-\r
- //merge the ranges of every process\r
- std::vector<std::pair<MPI_Offset, MPI_Offset>> ranges;\r
- for(int i=0; i<size; ++i)\r
- ranges.push_back(std::make_pair(min_offsets[i],max_offsets[i]));\r
- std::sort(ranges.begin(), ranges.end());\r
- std::vector<std::pair<MPI_Offset, MPI_Offset>> chunks;\r
- chunks.push_back(ranges[0]);\r
-\r
- unsigned int nchunks=0;\r
- unsigned int i=1;\r
- while(i < ranges.size()){\r
- if(ranges[i].second>chunks[nchunks].second){\r
- // else range included - ignore\r
- if(ranges[i].first>chunks[nchunks].second){\r
- //new disjoint range\r
- chunks.push_back(ranges[i]);\r
- nchunks++;\r
- } else {\r
- //merge ranges\r
- chunks[nchunks].second=ranges[i].second;\r
- }\r
- }\r
- i++;\r
- }\r
- //what do I need to read ?\r
- MPI_Offset totreads=0;\r
- for(i=0; i<chunks.size();i++){\r
- if(chunks[i].second < my_chunk_start)\r
- continue;\r
- else if (chunks[i].first > my_chunk_end)\r
- continue;\r
- else\r
- totreads += (std::min(chunks[i].second, my_chunk_end-1)-std::max(chunks[i].first, my_chunk_start));\r
- }\r
- XBT_DEBUG("will have to access %lld from my chunk", totreads);\r
-\r
- char* sendbuf= static_cast<char *>(smpi_get_tmp_sendbuffer(totreads));\r
-\r
- if(totreads>0){\r
- seek(min_offset, MPI_SEEK_SET);\r
- T(this,sendbuf,totreads/datatype->size(),datatype, status);\r
- }\r
- simgrid::smpi::Colls::alltoall(send_sizes, 1, MPI_INT, recv_sizes, 1, MPI_INT, comm_);\r
- int total_recv=0;\r
- for(int i=0;i<size;i++){\r
- recv_disps[i]=total_recv;\r
- total_recv+=recv_sizes[i];\r
- }\r
- //Set buf value to avoid copying dumb data\r
- simgrid::smpi::Colls::alltoallv(sendbuf, send_sizes, send_disps, MPI_BYTE,\r
- buf, recv_sizes, recv_disps, MPI_BYTE, comm_);\r
- status->count=count * datatype->size();\r
- smpi_free_tmp_buffer(sendbuf);\r
- xbt_free(send_sizes);\r
- xbt_free(recv_sizes);\r
- xbt_free(send_disps);\r
- xbt_free(recv_disps);\r
- xbt_free(min_offsets);\r
- xbt_free(max_offsets);\r
- return MPI_SUCCESS;\r
- }\r
-}\r
-}\r
-\r
-#endif\r
+/* Copyright (c) 2010-2019. The SimGrid Team.
+ * All rights reserved. */
+
+/* This program is free software; you can redistribute it and/or modify it
+ * under the terms of the license (GNU LGPL) which comes with this package. */
+
+#ifndef SMPI_FILE_HPP_INCLUDED
+#define SMPI_FILE_HPP_INCLUDED
+#include "simgrid/plugins/file_system.h"
+#include "smpi_comm.hpp"
+#include "smpi_coll.hpp"
+#include "smpi_datatype.hpp"
+#include "smpi_errhandler.hpp"
+#include "smpi_info.hpp"
+#include <algorithm>
+
+XBT_LOG_EXTERNAL_CATEGORY(smpi_pmpi);
+
+namespace simgrid{
+namespace smpi{
+class File{
+ MPI_Comm comm_;
+ int flags_;
+ simgrid::s4u::File* file_;
+ MPI_Info info_;
+ MPI_Offset* shared_file_pointer_;
+ s4u::MutexPtr shared_mutex_;
+ MPI_Win win_;
+ char* list_;
+ MPI_Errhandler errhandler_;
+
+ public:
+ File(MPI_Comm comm, const char *filename, int amode, MPI_Info info);
+ File(const File&) = delete;
+ File& operator=(const File&) = delete;
+ ~File();
+ int size();
+ int get_position(MPI_Offset* offset);
+ int get_position_shared(MPI_Offset* offset);
+ int flags();
+ MPI_Comm comm();
+ int sync();
+ int seek(MPI_Offset offset, int whence);
+ int seek_shared(MPI_Offset offset, int whence);
+ MPI_Info info();
+ void set_info( MPI_Info info);
+ static int read(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status);
+ static int read_shared(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status);
+ static int read_ordered(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status);
+ static int write(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status);
+ static int write_shared(MPI_File fh, const void *buf, int count,MPI_Datatype datatype, MPI_Status *status);
+ static int write_ordered(MPI_File fh, const void *buf, int count,MPI_Datatype datatype, MPI_Status *status);
+ template <int (*T)(MPI_File, void *, int, MPI_Datatype, MPI_Status *)> int op_all(void *buf, int count,MPI_Datatype datatype, MPI_Status *status);
+ static int close(MPI_File *fh);
+ static int del(const char *filename, MPI_Info info);
+ MPI_Errhandler errhandler();
+ void set_errhandler( MPI_Errhandler errhandler);
+};
+
+ /* Read_all, Write_all : loosely based on */
+ /* @article{Thakur:1996:ETM:245875.245879,*/
+ /* author = {Thakur, Rajeev and Choudhary, Alok},*/
+ /* title = {An Extended Two-phase Method for Accessing Sections of Out-of-core Arrays},*/
+ /* journal = {Sci. Program.},*/
+ /* issue_date = {Winter 1996},*/
+ /* pages = {301--317},*/
+ /* }*/
+ template <int (*T)(MPI_File, void *, int, MPI_Datatype, MPI_Status *)>
+ int File::op_all(void *buf, int count, MPI_Datatype datatype, MPI_Status *status){
+ //get min and max offsets from everyone.
+ int size = comm_->size();
+ int rank = comm_-> rank();
+ MPI_Offset min_offset = file_->tell();
+ MPI_Offset max_offset = (min_offset + count * datatype->size());//cheating, as we don't care about exact data location, we can skip extent
+ MPI_Offset* min_offsets = new MPI_Offset[size];
+ MPI_Offset* max_offsets = new MPI_Offset[size];
+ simgrid::smpi::Colls::allgather(&min_offset, 1, MPI_OFFSET, min_offsets, 1, MPI_OFFSET, comm_);
+ simgrid::smpi::Colls::allgather(&max_offset, 1, MPI_OFFSET, max_offsets, 1, MPI_OFFSET, comm_);
+ MPI_Offset min=min_offset;
+ MPI_Offset max=max_offset;
+ MPI_Offset tot= 0;
+ int empty=1;
+ for(int i=0;i<size;i++){
+ if(min_offsets[i]!=max_offsets[i])
+ empty=0;
+ tot+=(max_offsets[i]-min_offsets[i]);
+ if(min_offsets[i]<min)
+ min=min_offsets[i];
+ if(max_offsets[i]>max)
+ max=max_offsets[i];
+ }
+
+ XBT_CDEBUG(smpi_pmpi, "my offsets to read : %lld:%lld, global min and max %lld:%lld", min_offset, max_offset, min, max);
+ if(empty==1){
+ delete[] min_offsets;
+ delete[] max_offsets;
+ status->count=0;
+ return MPI_SUCCESS;
+ }
+ MPI_Offset total = max-min;
+ if(total==tot && (datatype->flags() & DT_FLAG_CONTIGUOUS)){
+ delete[] min_offsets;
+ delete[] max_offsets;
+ //contiguous. Just have each proc perform its read
+ status->count=count * datatype->size();
+ return T(this,buf,count,datatype, status);
+ }
+
+ //Interleaved case : How much do I need to read, and whom to send it ?
+ MPI_Offset my_chunk_start=(max-min+1)/size*rank;
+ MPI_Offset my_chunk_end=((max-min+1)/size*(rank+1));
+ XBT_CDEBUG(smpi_pmpi, "my chunks to read : %lld:%lld", my_chunk_start, my_chunk_end);
+ int* send_sizes = new int[size];
+ int* recv_sizes = new int[size];
+ int* send_disps = new int[size];
+ int* recv_disps = new int[size];
+ int total_sent=0;
+ for(int i=0;i<size;i++){
+ send_sizes[i]=0;
+ send_disps[i]=0;//cheat to avoid issues when send>recv as we use recv buffer
+ if((my_chunk_start>=min_offsets[i] && my_chunk_start < max_offsets[i])||
+ ((my_chunk_end<=max_offsets[i]) && my_chunk_end> min_offsets[i])){
+ send_sizes[i]=(std::min(max_offsets[i]-1, my_chunk_end-1)-std::max(min_offsets[i], my_chunk_start));
+ //store min and max offest to actually read
+ min_offset=std::min(min_offset, min_offsets[i]);
+ total_sent+=send_sizes[i];
+ XBT_CDEBUG(smpi_pmpi, "will have to send %d bytes to %d", send_sizes[i], i);
+ }
+ }
+ min_offset=std::max(min_offset, my_chunk_start);
+
+ //merge the ranges of every process
+ std::vector<std::pair<MPI_Offset, MPI_Offset>> ranges;
+ for(int i=0; i<size; ++i)
+ ranges.push_back(std::make_pair(min_offsets[i],max_offsets[i]));
+ std::sort(ranges.begin(), ranges.end());
+ std::vector<std::pair<MPI_Offset, MPI_Offset>> chunks;
+ chunks.push_back(ranges[0]);
+
+ unsigned int nchunks=0;
+ unsigned int i=1;
+ while(i < ranges.size()){
+ if(ranges[i].second>chunks[nchunks].second){
+ // else range included - ignore
+ if(ranges[i].first>chunks[nchunks].second){
+ //new disjoint range
+ chunks.push_back(ranges[i]);
+ nchunks++;
+ } else {
+ //merge ranges
+ chunks[nchunks].second=ranges[i].second;
+ }
+ }
+ i++;
+ }
+ //what do I need to read ?
+ MPI_Offset totreads=0;
+ for(i=0; i<chunks.size();i++){
+ if(chunks[i].second < my_chunk_start)
+ continue;
+ else if (chunks[i].first > my_chunk_end)
+ continue;
+ else
+ totreads += (std::min(chunks[i].second, my_chunk_end-1)-std::max(chunks[i].first, my_chunk_start));
+ }
+ XBT_CDEBUG(smpi_pmpi, "will have to access %lld from my chunk", totreads);
+
+ unsigned char* sendbuf = smpi_get_tmp_sendbuffer(total_sent);
+
+ if(totreads>0){
+ seek(min_offset, MPI_SEEK_SET);
+ T(this,sendbuf,totreads/datatype->size(),datatype, status);
+ }
+ simgrid::smpi::Colls::alltoall(send_sizes, 1, MPI_INT, recv_sizes, 1, MPI_INT, comm_);
+ int total_recv=0;
+ for(int i=0;i<size;i++){
+ recv_disps[i]=total_recv;
+ total_recv+=recv_sizes[i];
+ }
+ //Set buf value to avoid copying dumb data
+ simgrid::smpi::Colls::alltoallv(sendbuf, send_sizes, send_disps, MPI_BYTE,
+ buf, recv_sizes, recv_disps, MPI_BYTE, comm_);
+ status->count=count * datatype->size();
+ smpi_free_tmp_buffer(sendbuf);
+ delete[] send_sizes;
+ delete[] recv_sizes;
+ delete[] send_disps;
+ delete[] recv_disps;
+ delete[] min_offsets;
+ delete[] max_offsets;
+ return MPI_SUCCESS;
+ }
+}
+}
+
+#endif
