more optimization and a bug fix

ldmt-muri · Apr 23, 2015 · aba97d5 · aba97d5
1 parent 17d0d29
commit aba97d5
Show file tree

Hide file tree

Showing 4 changed files with 30 additions and 8 deletions.
diff --git a/alignment/LatentCrfAligner.cc b/alignment/LatentCrfAligner.cc
@@ -57,12 +57,17 @@ LatentCrfAligner::LatentCrfAligner(const string &textFilename,
 
   // encode the null token which is conventionally added to the beginning of the src sentnece. 
   NULL_TOKEN_STR = "__null__token__";
-  NULL_TOKEN = vocabEncoder.Encode(NULL_TOKEN_STR);
+  if (learningInfo.mpiWorld->rank() == 0) {
+    NULL_TOKEN = vocabEncoder.Encode(NULL_TOKEN_STR);
+  }
+  boost::mpi::broadcast<int64_t>(*learningInfo.mpiWorld, NULL_TOKEN, 0);
   assert(NULL_TOKEN != vocabEncoder.UnkInt());
 
   // read and encode tgt words and their classes (e.g. brown clusters)
-  EncodeTgtWordClasses();
-
+  if (learningInfo.mpiWorld->rank() == 0) {
+    EncodeTgtWordClasses();
+  }
+
   // read and encode data
   srcSents.clear();
   tgtSents.clear();

diff --git a/alignment/LatentCrfAligner.h b/alignment/LatentCrfAligner.h
@@ -2,6 +2,15 @@
 #define _LATENT_CRF_ALIGNER_H_
 
 #include <fstream>
+
+#include "mpi.h"
+
+#include <boost/mpi/environment.hpp>
+#include <boost/mpi/communicator.hpp>
+#include <boost/mpi/nonblocking.hpp>
+#include <boost/mpi/collectives.hpp>
+
+
 #include "../core/LatentCrfModel.h"
 
 class LatentCrfAligner : public LatentCrfModel {

diff --git a/core/LearningInfo.h b/core/LearningInfo.h
@@ -193,7 +193,7 @@ class LearningInfo {
 
   // should be called only once in the constructor
   void SetSharedMemorySegment(bool create) {
-    size_t segmentSize = 30 * 1024; // in GBs
+    size_t segmentSize = 256 * 1024; // in GBs
     segmentSize *= 1024 * 1024;
     string SEGMENT_NAME = outputFilenamePrefix + ".segment";
     using namespace boost::interprocess;

diff --git a/core/VocabEncoder.h b/core/VocabEncoder.h
@@ -296,13 +296,17 @@ class VocabEncoder {
 
       if (end_of_file) {
 	break;
-      } else {
+      }
+
+      // Wait till the master has encoded this line.
+      if (learningInfo.mpiWorld->rank() != 0) {
 	boost::mpi::broadcast<std::string>(*learningInfo.mpiWorld, line, 0);
       }
 
       // skip empty lines
       if(line.size() == 0) {
-	continue;
+	cerr << "Blank lines are not allowed in parallel data. Will die." << endl;
+	exit(1);
       }
       lineNumber++;
 
@@ -314,10 +318,14 @@ class VocabEncoder {
       srcSents.resize(lineNumber+1);
       tgtSents.resize(lineNumber+1);
       vector<int64_t> temp;
-
       Encode(splits, temp);
-
       assert(splits.size() == temp.size());
+
+      // Having encoded the words in this line, the master now sends this line to the slaves.
+      if (learningInfo.mpiWorld->rank() == 0) {
+	boost::mpi::broadcast<std::string>(*learningInfo.mpiWorld, line, 0);
+      }
+
       // src sent is written before tgt sent
       bool src = true;
       if(nullToken.size() > 0) {