From fac397eda7e3352d99b85d09050f496d8b9e06eb Mon Sep 17 00:00:00 2001
From: Dongryeol Lee <drselee@gmail.com>
Date: Fri, 21 Jan 2011 04:45:05 +0000
Subject: [PATCH] Need to write a non-sampling based binary tree builder.

---
 .../core/gnp/distributed_dualtree_dfs.h       | 22 ++++++++++++++++++
 .../core/gnp/distributed_dualtree_dfs_dev.h   | 23 +++++++++++++++----
 .../core/parallel/distributed_tree_builder.h  |  2 +-
 3 files changed, 42 insertions(+), 5 deletions(-)

diff --git a/fastlib/trunk/contrib/dongryel/thesis_research/core/gnp/distributed_dualtree_dfs.h b/fastlib/trunk/contrib/dongryel/thesis_research/core/gnp/distributed_dualtree_dfs.h
index b652129539..92adf0ca00 100644
--- a/fastlib/trunk/contrib/dongryel/thesis_research/core/gnp/distributed_dualtree_dfs.h
+++ b/fastlib/trunk/contrib/dongryel/thesis_research/core/gnp/distributed_dualtree_dfs.h
@@ -9,6 +9,7 @@
 #define CORE_GNP_DISTRIBUTED_DUALTREE_DFS_H
 
 #include <boost/mpi/communicator.hpp>
+#include <boost/thread.hpp>
 #include <boost/tuple/tuple.hpp>
 #include "core/math/range.h"
 
@@ -34,14 +35,30 @@ class DistributedDualtreeDfs {
 
   private:
 
+    /** @brief The pointer to the boost communicator.
+     */
     boost::mpi::communicator *world_;
 
+    /** @brief The problem definition for the distributed computation.
+     */
     DistributedProblemType *problem_;
 
+    /** @brief The distributed query table.
+     */
     DistributedTableType *query_table_;
 
+    /** @brief The distributed reference table.
+     */
     DistributedTableType *reference_table_;
 
+    /** @brief The worker pool.
+     */
+    boost::thread_group worker_pool_;
+
+    /** @brief The work queue from which the workers work.
+     */
+    std::multimap<TreeType *, TreeType *> work_queue_;
+
   private:
 
     template<typename MetricType>
@@ -63,6 +80,11 @@ class DistributedDualtreeDfs {
       const MetricType &metric,
       TreeType *qnode, ResultType *query_results);
 
+    static void DoIt_();
+
+    /** @brief The class used for prioritizing a computation object
+     *         (query, reference pair).
+     */
     class PrioritizeTasks_:
       public std::binary_function <
         FrontierObjectType &, FrontierObjectType &, bool > {
diff --git a/fastlib/trunk/contrib/dongryel/thesis_research/core/gnp/distributed_dualtree_dfs_dev.h b/fastlib/trunk/contrib/dongryel/thesis_research/core/gnp/distributed_dualtree_dfs_dev.h
index 556466ba91..ed3482b7b1 100644
--- a/fastlib/trunk/contrib/dongryel/thesis_research/core/gnp/distributed_dualtree_dfs_dev.h
+++ b/fastlib/trunk/contrib/dongryel/thesis_research/core/gnp/distributed_dualtree_dfs_dev.h
@@ -13,6 +13,7 @@
 #include <boost/mpi.hpp>
 #include <boost/tuple/tuple.hpp>
 #include <list>
+#include <map>
 #include <queue>
 #include "core/gnp/distributed_dualtree_dfs.h"
 #include "core/gnp/dualtree_dfs_dev.h"
@@ -30,6 +31,13 @@ extern core::table::MemoryMappedFile *global_m_file_;
 namespace core {
 namespace gnp {
 
+template<typename DistributedProblemType>
+void DistributedDualtreeDfs<DistributedProblemType>::DoIt_() {
+
+  // Dequeue a work.
+
+}
+
 template<typename DistributedProblemType>
 template<typename MetricType>
 void DistributedDualtreeDfs<DistributedProblemType>::AllToAllReduce_(
@@ -214,6 +222,13 @@ void DistributedDualtreeDfs<DistributedProblemType>::Init(
   if(query_table_ != reference_table_) {
     ResetStatisticRecursion_(reference_table_->get_tree(), reference_table_);
   }
+
+  // Start the thread worker pool.
+  //const int num_worker_threads = 8;
+  //for(int i = 0; i < num_worker_threads; i++) {
+  //worker_pool_.create_thread(
+  //  DistributedDualtreeDfs < DistributedProblemType >::DoIt_);
+  //}
 }
 
 template<typename DistributedProblemType>
@@ -236,11 +251,11 @@ void DistributedDualtreeDfs<DistributedProblemType>::Compute(
   PreProcessReferenceTree_(reference_table_->get_tree());
   PreProcessReferenceTree_(reference_table_->local_table()->get_tree());
 
-  // Figure out each process's work using the global tree. This is
-  // done using a naive approach where the global goal is to complete
-  // a 2D matrix workspace. This is currently doing an all-reduce type
-  // of exchange.
+  // Figure out each process's work using the global tree. a 2D matrix
+  // workspace. This is currently doing an all-reduce type of
+  // exchange.
   AllToAllReduce_(metric, query_results);
+  worker_pool_.join_all();
   world_->barrier();
 
   // Postprocess.
diff --git a/fastlib/trunk/contrib/dongryel/thesis_research/core/parallel/distributed_tree_builder.h b/fastlib/trunk/contrib/dongryel/thesis_research/core/parallel/distributed_tree_builder.h
index e8c10f47cb..51cfd7f1a0 100644
--- a/fastlib/trunk/contrib/dongryel/thesis_research/core/parallel/distributed_tree_builder.h
+++ b/fastlib/trunk/contrib/dongryel/thesis_research/core/parallel/distributed_tree_builder.h
@@ -541,7 +541,7 @@ class DistributedTreeBuilder {
         global_root_bound.get(i) = global_root_bound_vector[i];
       }
 
-      for(int num_outer_it = 0; num_outer_it < 3; num_outer_it++) {
+      for(int num_outer_it = 0; num_outer_it < 30; num_outer_it++) {
 
         // Build the initial sample tree.
         std::vector<TreeType *> top_leaf_nodes;