Learning based LB changes

CMakeLists.txt

@@ -779,7 +779,7 @@ set(src-util-h-sources src/util/SSE-Double.h src/util/SSE-Float.h
     src/util/treeStrategy_nodeAware_minBytes.h
     src/util/treeStrategy_nodeAware_minGens.h src/util/treeStrategy_topoUnaware.h
     src/util/uFcontext.h src/util/uJcontext.h src/util/valgrind.h
-    src/util/vector2d.h)
+    src/util/vector2d.h src/util/fastforest.h)
 
 foreach(filename ${src-util-h-sources})
     configure_file(${filename} include/ COPYONLY)

cmake/converse.cmake

@@ -155,6 +155,7 @@ set(conv-util-cxx-sources
     src/util/pup_util.C
     src/util/pup_xlater.C
     src/util/spanningTree.C
+        src/util/fastforest.C
 )
 
 if(CMK_CAN_LINK_FORTRAN)

src/ck-ldb/LBManager.C

@@ -198,6 +198,8 @@ void _loadbalancerInit()
   CmiGetArgStringDesc(argv, "+MetaLBModelDir", &_lb_args.metaLbModelDir(),
                       "Use this directory to read model for MetaLB");
 
+  _lb_args.treeMetaLbOn() = CmiGetArgFlagDesc(argv, "+TreeMetaLB", "use MetaLB within TreeLB");
+
   if (_lb_args.metaLbOn() && _lb_args.metaLbModelDir() != nullptr)
   {
 #if CMK_USE_ZLIB

src/ck-ldb/LBManager.h

@@ -40,6 +40,7 @@ class CkLBArgs
   double _lb_targetRatio;  // Specifies the target load ratio for LBs that aim for a
                            // particular load ratio
   bool _lb_metaLbOn;
+  bool _lb_treeMetaLbOn;
   char* _lb_metaLbModelDir;
   char* _lb_treeLBFile = (char*)"treelb.json";
 
@@ -56,6 +57,7 @@ class CkLBArgs
     _lb_maxDistPhases = 10;
     _lb_targetRatio = 1.05;
     _lb_metaLbOn = false;
+    _lb_treeMetaLbOn = false;
     _lb_metaLbModelDir = nullptr;
   }
   inline char*& treeLBFile() { return _lb_treeLBFile; }
@@ -78,6 +80,7 @@ class CkLBArgs
   inline int& maxDistPhases() { return _lb_maxDistPhases; }
   inline double& targetRatio() { return _lb_targetRatio; }
   inline bool& metaLbOn() { return _lb_metaLbOn; }
+  inline bool& treeMetaLbOn() {return _lb_treeMetaLbOn; }
   inline char*& metaLbModelDir() { return _lb_metaLbModelDir; }
 };
 

src/ck-ldb/MetaBalancer.C

@@ -23,6 +23,7 @@
 #define NEGLECT_IDLE 2 // Should never be == 1
 #define MIN_STATS 6
 #define STATS_COUNT 29 // The number of stats collected during reduction
+#define CLASSES 6
 
 #define MAXDOUBLE  std::numeric_limits<double>::max()
 
@@ -36,6 +37,7 @@ using std::max;
 CkReductionMsg* lbDataCollection(int nMsg, CkReductionMsg** msgs) {
   double *lb_data;
   lb_data = (double*)msgs[0]->getData();
+  lb_data = (double*)msgs[0]->getData();
   for (int i = 1; i < nMsg; i++) {
     CkAssert(msgs[i]->getSize() == STATS_COUNT*sizeof(double));
     if (msgs[i]->getSize() != STATS_COUNT*sizeof(double)) {
@@ -178,6 +180,12 @@ void MetaBalancer::init(void) {
       srand(time(NULL));
       rFmodel = new ForestModel;
       rFmodel->readModel(_lb_args.metaLbModelDir());
+
+      std::vector <std::string> features{"f0", "f1", "f2", "f3", "f4", "f5", "f6", "f7", "f8", "f9", "f10", "f11",
+                                         "f12", "f13", "f14", "f15", "f16", "f17", "f18", "f19", "f20", "f21", "f22",
+                                         "f23", "f24"};
+      xgboost = fastforest::load_txt("model/model.txt", features, CLASSES);
+
     }
   }
 }
@@ -295,9 +303,9 @@ bool MetaBalancer::AddLoad(int it_n, double load) {
   int index = it_n % VEC_SIZE;
   total_count_vec[index]++;
   adaptive_struct.total_syncs_called++;
-  CkPrintf("At PE %d Total contribution for iteration %d is %d \
+  DEBAD(("At PE %d Total contribution for iteration %d is %d \
       total objs %d\n", CkMyPe(), it_n, total_count_vec[index],
-      lbmanager->GetObjDataSz());
+      lbmanager->GetObjDataSz()));
 
   if (it_n <= adaptive_struct.finished_iteration_no) {
     CkAbort("Error!! Received load for iteration that has contributed\n");
@@ -324,6 +332,7 @@ bool MetaBalancer::AddLoad(int it_n, double load) {
   return true;
 }
 
+//NOTE: Data Collection
 void MetaBalancer::ContributeStats(int it_n) {
 #if CMK_LBDB_ON
   int index = it_n % VEC_SIZE;
@@ -493,7 +502,7 @@ void MetaBalancer::ReceiveMinStats(double *load, int n) {
 //        avg_hops, avg_hop_bytes, _lb_args.alpha(), _lb_args.beta(),
 //        app_iteration_time);
 
-    DEBAD(
+   DEBAD(
         ("Features:%lf %lf %lf %lf %lf %lf %lf %lf \
        %lf %lf %lf %lf %lf %lf %lf %lf %lf \
        %lf %lf %lf %lf %lf %lf %lf %d %lf\n",
@@ -532,10 +541,24 @@ void MetaBalancer::ReceiveMinStats(double *load, int n) {
                                     avg_hops,
                                     avg_hop_Kbytes,
                                     comm_comp_ratio};
+
+      //Note:Predict LB
       // Model returns value [1,num_lbs]
       int predicted_lb = rFmodel->forestTest(test_data, 1, 26);
       DEBAD(("***********Final classification = %d *****************\n", predicted_lb));
 
+      std::vector<double> prob = xgboost.softmax(test_data.data());
+      int cur_pred = 1;
+      double cur_prob = 0.0;
+      for(int i = 0; i < prob.size(); ++i) {
+        if(prob[i] > cur_prob) {
+          cur_prob = prob[i];
+          cur_pred = i+1;
+        }
+      }
+
+      predicted_lb = cur_pred;
+
       // predicted_lb-1 since predicted_lb class count in the model starts at 1
       thisProxy.MetaLBSetLBOnChares(current_balancer, predicted_lb - 1);
       current_balancer = predicted_lb - 1;

src/ck-ldb/MetaBalancer.h

@@ -32,6 +32,7 @@
 
 #include "LBManager.h"
 #include "RandomForestModel.h"
+#include "fastforest.h"
 #include <vector>
 
 #include "MetaBalancer.decl.h"
@@ -89,7 +90,9 @@ class MetaBalancer : public CBase_MetaBalancer {
  MetaBalancer(void) : rFmodel(NULL) { init(); }
  MetaBalancer(CkMigrateMessage* m) : CBase_MetaBalancer(m) { init(); }
  ~MetaBalancer() {
-   if (CkMyPe() == 0) delete rFmodel;
+   if (CkMyPe() == 0) {
+     delete rFmodel;
+   }
  }
 
 private:
@@ -176,6 +179,7 @@ class MetaBalancer : public CBase_MetaBalancer {
   int total_ovld_pes;
   int current_balancer;
   ForestModel* rFmodel;
+  fastforest::FastForest xgboost;
 
   struct AdaptiveData {
     double iteration;

src/ck-ldb/TreeLB.C

@@ -136,6 +136,7 @@ TreeLB::~TreeLB()
 #endif
 }
 
+//Note: Each level configured with choice of load balancer
 void TreeLB::configure(LBTreeBuilder& builder, json& config)
 {
 #if CMK_LBDB_ON

src/ck-ldb/TreeLB.h

@@ -6,12 +6,18 @@
 #include "BaseLB.h"
 #include "TreeLB.decl.h"
 #include "json.hpp"
+#include "fastforest.h"
 #include <vector>
 using json = nlohmann::json;
+//using namespace rfmodel;
 
 #define DEBUG__TREE_LB_L1 0
 #define DEBUG__TREE_LB_L2 0
 #define DEBUG__TREE_LB_L3 0
+#define STATS_COUNT 29
+#define CLASSES 3
+
+extern CkLBArgs _lb_args;
 
 void CreateTreeLB();
 
@@ -24,12 +30,28 @@ class TreeLBMessage
 {
  public:
   uint8_t level;
+  //TODO Metabalancer stats
+  double lb_data[STATS_COUNT];
   // WARNING: don't add any virtual methods here
 };
 
 class LevelLogic
 {
  public:
+
+    //TODO: Initialize if only Metabalancer called
+    LevelLogic() {
+      //rfModel = new ForestModel;
+      //rfmodel->readModel(_lb_args.metaLbModelDir());
+      if(_lb_args.treeMetaLbOn()) {
+        std::vector <std::string> features{"f0", "f1", "f2", "f3", "f4", "f5", "f6", "f7", "f8", "f9", "f10", "f11",
+                                           "f12", "f13", "f14", "f15", "f16", "f17", "f18", "f19", "f20", "f21", "f22",
+                                           "f23", "f24"};
+        xgboost = fastforest::load_txt("model_tree/model.txt", features, CLASSES);
+      }
+    }
+
+
   virtual ~LevelLogic() {}
 
   /// return msg with lb stats for this PE. only needed at leaves
@@ -125,6 +147,9 @@ class LevelLogic
 
  protected:
   std::vector<TreeLBMessage*> stats_msgs;
+  //TODO Meta model
+  //ForestModel* rfmodel;
+  fastforest::FastForest xgboost;
 };
 
 class LBTreeBuilder;
@@ -140,6 +165,7 @@ class TreeLB : public CBase_TreeLB
     loadConfigFile(opts);
     init(opts);
   }
+
   TreeLB(CkMigrateMessage* m) : CBase_TreeLB(m) {}
   virtual ~TreeLB();