Begin working on the communication optimization for partial shared malloc.

diff --git a/include/smpi/smpi.h b/include/smpi/smpi.h
index 7f33224..b8a1475 100644 (file)
--- a/include/smpi/smpi.h
+++ b/include/smpi/smpi.h
@@ -872,7 +872,6 @@ XBT_PUBLIC(void) smpi_trace_set_call_location__(const char *file, int* line);
 #define SMPI_SAMPLE_DELAY(duration) for(smpi_execute(duration); 0; )
 #define SMPI_SAMPLE_FLOPS(flops) for(smpi_execute_flops(flops); 0; )
 
-XBT_PUBLIC(int) smpi_is_shared(void *buf);
 XBT_PUBLIC(void *) smpi_shared_malloc(size_t size, const char *file, int line);
 #define SMPI_SHARED_MALLOC(size) smpi_shared_malloc(size, __FILE__, __LINE__)
 XBT_PUBLIC(void *) smpi_shared_malloc_global__(size_t size, const char *file, int line, int *shared_block_offsets, int nb_shared_blocks);

diff --git a/include/smpi/smpi_shared_malloc.hpp b/include/smpi/smpi_shared_malloc.hpp
new file mode 100644 (file)
index 0000000..cf554ae
--- /dev/null
+++ b/include/smpi/smpi_shared_malloc.hpp
@@ -0,0 +1,16 @@
+#ifndef SMPI_SHARED_HPP
+#define SMPI_SHARED_HPP
+#include <xbt/function_types.h>
+#include <xbt/misc.h>
+#include <vector>
+
+
+/*
+ * We cannot put this declaration in smpi.h, since we use C++ features.
+ */
+
+
+XBT_PUBLIC(int) smpi_is_shared(void* ptr, std::vector<std::pair<int, int>> &private_blocks);
+
+
+#endif

diff --git a/src/smpi/smpi_datatype.cpp b/src/smpi/smpi_datatype.cpp
index 52ca497..3b36cab 100644 (file)
--- a/src/smpi/smpi_datatype.cpp
+++ b/src/smpi/smpi_datatype.cpp
@@ -271,11 +271,14 @@ int Datatype::copy(void *sendbuf, int sendcount, MPI_Datatype sendtype,
                        void *recvbuf, int recvcount, MPI_Datatype recvtype){
   int count;
 
+// FIXME Handle the case of a partial shared malloc.
+#if 0
   if(smpi_is_shared(sendbuf)){
     XBT_DEBUG("Copy input buf %p is shared. Let's ignore it.", sendbuf);
   }else if(smpi_is_shared(recvbuf)){
     XBT_DEBUG("Copy output buf %p is shared. Let's ignore it.", recvbuf);
   }
+#endif
 
   if(smpi_privatize_global_variables){
     smpi_switch_data_segment(smpi_process()->index());

diff --git a/src/smpi/smpi_global.cpp b/src/smpi/smpi_global.cpp
index 0acc0d4..a343824 100644 (file)
--- a/src/smpi/smpi_global.cpp
+++ b/src/smpi/smpi_global.cpp
@@ -7,6 +7,7 @@
 #include "private.h"
 #include "private.hpp"
 #include "simgrid/s4u/Mailbox.hpp"
+#include "smpi/smpi_shared_malloc.hpp"
 #include "simgrid/sg_config.h"
 #include "src/kernel/activity/SynchroComm.hpp"
 #include "src/mc/mc_record.h"
@@ -103,17 +104,47 @@ void smpi_comm_set_copy_data_callback(void (*callback) (smx_activity_t, void*, s
   smpi_comm_copy_data_callback = callback;
 }
 
+std::vector<std::pair<int, int>> merge_private_blocks(std::vector<std::pair<int, int>> src, std::vector<std::pair<int, int>> dst) {
+  std::vector<std::pair<int, int>> result;
+  int i_src=0, i_dst=0;
+  while(i_src < src.size() && i_dst < dst.size()) {
+    std::pair<int, int> block;
+    if(src[i_src].first < dst[i_dst].first) {
+      block = src[i_src];
+      i_src ++;
+    }
+    else {
+      block = dst[i_dst];
+      i_dst ++;
+    }
+    if(block.first <= result.back().second) { // overlapping with the last block inserted
+      result.back().second = std::max(result.back().second, block.second);
+    }
+    else { // not overlapping, we insert a new block
+      result.push_back(block);
+    }
+  }
+  for(; i_src < src.size(); i_src++) {
+    result.push_back(src[i_src]);
+  }
+  for(; i_dst < dst.size(); i_dst++) {
+    result.push_back(dst[i_dst]);
+  }
+  return result;
+}
+
 void smpi_comm_copy_buffer_callback(smx_activity_t synchro, void *buff, size_t buff_size)
 {
-
   simgrid::kernel::activity::Comm *comm = dynamic_cast<simgrid::kernel::activity::Comm*>(synchro);
-
+  int src_shared=0, dst_shared=0;
+  std::vector<std::pair<int, int>> src_private_blocks;
+  std::vector<std::pair<int, int>> dst_private_blocks;
   XBT_DEBUG("Copy the data over");
-  if(smpi_is_shared(buff)){
+  if(src_shared=smpi_is_shared(buff, src_private_blocks))
     XBT_DEBUG("Sender %p is shared. Let's ignore it.", buff);
-  }else if(smpi_is_shared((char*)comm->dst_buff)){
-    XBT_DEBUG("Receiver %p is shared. Let's ignore it.", (char*)comm->dst_buff);
-  }else{
+  if(dst_shared=smpi_is_shared((char*)comm->dst_buff, src_private_blocks))
+    XBT_DEBUG("Receiver %p is shared. Let's ignore it.", (char*)comm->dst_buff); 
+  if(!src_shared && !dst_shared){
     void* tmpbuff=buff;
     if((smpi_privatize_global_variables) && (static_cast<char*>(buff) >= smpi_start_data_exe)
         && (static_cast<char*>(buff) < smpi_start_data_exe + smpi_size_data_exe )

diff --git a/src/smpi/smpi_request.cpp b/src/smpi/smpi_request.cpp
index fd91cf4..b3988d8 100644 (file)
--- a/src/smpi/smpi_request.cpp
+++ b/src/smpi/smpi_request.cpp
@@ -108,7 +108,8 @@ namespace smpi{
 Request::Request(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm, unsigned flags) : buf_(buf), old_type_(datatype), src_(src), dst_(dst), tag_(tag), comm_(comm), flags_(flags) 
 {
   void *old_buf = nullptr;
-  if(((((flags & RECV) != 0) && ((flags & ACCUMULATE) !=0)) || (datatype->flags() & DT_FLAG_DERIVED)) && (!smpi_is_shared(buf_))){
+// FIXME Handle the case of a partial shared malloc.
+  if(((((flags & RECV) != 0) && ((flags & ACCUMULATE) !=0)) || (datatype->flags() & DT_FLAG_DERIVED))) { // && (!smpi_is_shared(buf_))){
     // This part handles the problem of non-contiguous memory
     old_buf = buf;
     buf_ = count==0 ? nullptr : xbt_malloc(count*datatype->size());
@@ -758,7 +759,8 @@ void Request::finish_wait(MPI_Request* request, MPI_Status * status)
     req->print_request("Finishing");
     MPI_Datatype datatype = req->old_type_;
 
-    if((((req->flags_ & ACCUMULATE) != 0) || (datatype->flags() & DT_FLAG_DERIVED)) && (!smpi_is_shared(req->old_buf_))){
+// FIXME Handle the case of a partial shared malloc.
+    if((((req->flags_ & ACCUMULATE) != 0) || (datatype->flags() & DT_FLAG_DERIVED))){// && (!smpi_is_shared(req->old_buf_))){
 
       if (!smpi_process()->replaying()){
         if( smpi_privatize_global_variables != 0 && (static_cast<char*>(req->old_buf_) >= smpi_start_data_exe)

diff --git a/src/smpi/smpi_shared.cpp b/src/smpi/smpi_shared.cpp
index c5208ce..731f4ae 100644 (file)
--- a/src/smpi/smpi_shared.cpp
+++ b/src/smpi/smpi_shared.cpp
@@ -114,6 +114,7 @@ typedef std::unordered_map<smpi_source_location, shared_data_t>::value_type shar
 
 typedef struct {
   size_t size;
+  std::vector<std::pair<int, int>> private_blocks;
   shared_data_key_type* data;
 } shared_metadata_t;
 
@@ -240,6 +241,9 @@ void *smpi_shared_malloc_global__(size_t size, const char *file, int line, int *
     xbt_assert(0 <= start_offset,          "start_offset (%d) should be greater than 0", start_offset);
     xbt_assert(start_offset < stop_offset, "start_offset (%d) should be lower than stop offset (%d)", start_offset, stop_offset);
     xbt_assert(stop_offset <= size,         "stop_offset (%d) should be lower than size (%lu)", stop_offset, size);
+    if(i_block < nb_shared_blocks-1)
+      xbt_assert(stop_offset < shared_block_offsets[2*i_block+2],
+              "stop_offset (%d) should be lower than its successor start offset (%d)", stop_offset, shared_block_offsets[2*i_block+2]);
 //    fprintf(stderr, "shared block 0x%x - 0x%x\n", start_offset, stop_offset);
     int start_block_offset = ALIGN_UP(start_offset, smpi_shared_malloc_blocksize);
     int stop_block_offset = ALIGN_DOWN(stop_offset, smpi_shared_malloc_blocksize);
@@ -281,16 +285,23 @@ void *smpi_shared_malloc_global__(size_t size, const char *file, int line, int *
     }
   }
 
-  if(nb_shared_blocks == 1 && shared_block_offsets[0] == 0 && shared_block_offsets[1] == size) {
-    shared_metadata_t newmeta;
-    //register metadata for memcpy avoidance
-    shared_data_key_type* data = (shared_data_key_type*)xbt_malloc(sizeof(shared_data_key_type));
-    data->second.fd = -1;
-    data->second.count = 1;
-    newmeta.size = size;
-    newmeta.data = data;
-    allocs_metadata[mem] = newmeta;
+  shared_metadata_t newmeta;
+  //register metadata for memcpy avoidance
+  shared_data_key_type* data = (shared_data_key_type*)xbt_malloc(sizeof(shared_data_key_type));
+  data->second.fd = -1;
+  data->second.count = 1;
+  newmeta.size = size;
+  newmeta.data = data;
+  if(shared_block_offsets[0] > 0) {
+    newmeta.private_blocks.push_back(std::make_pair(0, shared_block_offsets[0]));
+  }
+  for(int i_block = 0; i_block < nb_shared_blocks-1; i_block ++) {
+    newmeta.private_blocks.push_back(std::make_pair(shared_block_offsets[2*i_block+1], shared_block_offsets[2*i_block+2]));
+  }
+  if(shared_block_offsets[nb_shared_blocks-1] < size) {
+    newmeta.private_blocks.push_back(std::make_pair(shared_block_offsets[nb_shared_blocks-1], size));
   }
+  allocs_metadata[mem] = newmeta;
 
   return mem;
 }
@@ -326,18 +337,23 @@ void *smpi_shared_malloc(size_t size, const char *file, int line) {
   return mem;
 }
 
-int smpi_is_shared(void* ptr){
+int smpi_is_shared(void* ptr, std::vector<std::pair<int, int>> &private_blocks){
+  private_blocks.clear(); // being paranoid
   if (allocs_metadata.empty())
     return 0;
   if ( smpi_cfg_shared_malloc == shmalloc_local || smpi_cfg_shared_malloc == shmalloc_global) {
     auto low = allocs_metadata.lower_bound(ptr);
-    if (low->first==ptr)
+    if (low->first==ptr) {
+      private_blocks = low->second.private_blocks;
       return 1;
+    }
     if (low == allocs_metadata.begin())
       return 0;
     low --;
-    if (ptr < (char*)low->first + low->second.size)
+    if (ptr < (char*)low->first + low->second.size) {
+      private_blocks = low->second.private_blocks;
       return 1;
+    }
     return 0;
   } else {
     return 0;