Working copy optimization \o/

[simgrid.git] / src / smpi / smpi_global.cpp

diff --git a/src/smpi/smpi_global.cpp b/src/smpi/smpi_global.cpp
index db497bf..1ad2a30 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"
@@ -103,12 +104,45 @@ void smpi_comm_set_copy_data_callback(void (*callback) (smx_activity_t, void*, s
   smpi_comm_copy_data_callback = callback;
 }
 
+void print(std::vector<std::pair<int, int>> vec) {
+    fprintf(stderr, "{");
+    for(auto elt: vec) {
+        fprintf(stderr, "(0x%x, 0x%x),", elt.first, elt.second);
+    }
+    stderr, fprintf(stderr, "}\n");
+}
+void memcpy_private(void *dest, const void *src, size_t n, std::vector<std::pair<int, int>> &private_blocks) {
+  for(auto block : private_blocks) {
+    memcpy((uint8_t*)dest+block.first, (uint8_t*)src+block.first, block.second-block.first);
+  }
+}
+
 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;
+  int src_offset, dst_offset;
+  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(src_shared=smpi_is_shared(buff, src_private_blocks, &src_offset)) {
+    XBT_DEBUG("Sender %p is shared. Let's ignore it.", buff);
+    src_private_blocks = shift_private_blocks(src_private_blocks, src_offset);
+  }
+  else {
+    src_private_blocks.clear();
+    src_private_blocks.push_back(std::make_pair(0, buff_size));
+  }
+  if(dst_shared=smpi_is_shared((char*)comm->dst_buff, dst_private_blocks, &dst_offset)) {
+    XBT_DEBUG("Receiver %p is shared. Let's ignore it.", (char*)comm->dst_buff);
+    dst_private_blocks = shift_private_blocks(dst_private_blocks, dst_offset);
+  }
+  else {
+    dst_private_blocks.clear();
+    dst_private_blocks.push_back(std::make_pair(0, buff_size));
+  }
+  auto private_blocks = merge_private_blocks(src_private_blocks, dst_private_blocks);
   void* tmpbuff=buff;
-  simgrid::kernel::activity::Comm *comm = dynamic_cast<simgrid::kernel::activity::Comm*>(synchro);
-
   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 )
     ){
@@ -117,7 +151,7 @@ void smpi_comm_copy_buffer_callback(smx_activity_t synchro, void *buff, size_t b
        smpi_switch_data_segment(
            (static_cast<simgrid::smpi::Process*>((static_cast<simgrid::MsgActorExt*>(comm->src_proc->data)->data))->index()));
        tmpbuff = static_cast<void*>(xbt_malloc(buff_size));
-       memcpy(tmpbuff, buff, buff_size);
+       memcpy_private(tmpbuff, buff, buff_size, private_blocks);
   }
 
   if((smpi_privatize_global_variables) && ((char*)comm->dst_buff >= smpi_start_data_exe)
@@ -128,7 +162,8 @@ void smpi_comm_copy_buffer_callback(smx_activity_t synchro, void *buff, size_t b
   }
 
   XBT_DEBUG("Copying %zu bytes from %p to %p", buff_size, tmpbuff,comm->dst_buff);
-  memcpy(comm->dst_buff, tmpbuff, buff_size);
+  memcpy_private(comm->dst_buff, tmpbuff, buff_size, private_blocks);
+
   if (comm->detached) {
     // if this is a detached send, the source buffer was duplicated by SMPI
     // sender to make the original buffer available to the application ASAP
@@ -137,8 +172,8 @@ void smpi_comm_copy_buffer_callback(smx_activity_t synchro, void *buff, size_t b
     //xbt_free(comm->comm.src_data);// inside SMPI the request is kept inside the user data and should be free
     comm->src_buff = nullptr;
   }
-
   if(tmpbuff!=buff)xbt_free(tmpbuff);
+
 }
 
 void smpi_comm_null_copy_buffer_callback(smx_activity_t comm, void *buff, size_t buff_size)
@@ -297,6 +332,7 @@ void smpi_global_destroy()
   int count = smpi_process_count();
 
   smpi_bench_destroy();
+  smpi_shared_destroy();
   if (MPI_COMM_WORLD != MPI_COMM_UNINITIALIZED){
       delete MPI_COMM_WORLD->group();
       MSG_barrier_destroy(process_data[0]->finalization_barrier());
@@ -391,6 +427,7 @@ static void smpi_init_logs(){
   XBT_LOG_CONNECT(smpi_request);
   XBT_LOG_CONNECT(smpi_replay);
   XBT_LOG_CONNECT(smpi_rma);
+  XBT_LOG_CONNECT(smpi_shared);
   XBT_LOG_CONNECT(smpi_utils);
 }
 }