Changeset 2236

Timestamp:

09/20/21 18:18:17 (3 years ago)

Author:

ymipsl

Message:

Fix problem in remoteConnector when computing grid to sent to server.
Some optimisations when grid is not distributed need knowledge of the workflow view.
New CGridClientServerConnector class created based on CGridRemoteConnector.

YM

Location:

XIOS/dev/dev_ym/XIOS_COUPLING/src

Files:

• distribution/grid_client_server_remote_connector.cpp (added)
• distribution/grid_client_server_remote_connector.hpp (added)
• distribution/grid_remote_connector.cpp (modified) (6 diffs)
• distribution/grid_remote_connector.hpp (modified) (2 diffs)
• node/grid.cpp (modified) (5 diffs)
• test/generic_testcase.f90 (modified) (2 diffs)

XIOS/dev/dev_ym/XIOS_COUPLING/src/distribution/grid_remote_connector.cpp

@@ -111 +111 @@
 /**
   * \brief Compute the connector, i.e. compute the \b elements_ attribute. 
+  * \detail Depending of the distributions of the view computed in the computeViewDistribution() call, the connector is computed in computeConnectorMethods(), and to achieve better optimisation
+  *         some redondant ranks can be removed from the elements_ map.
+  */
+  void CGridRemoteConnector::computeConnector(void)
+  {
+    computeViewDistribution() ;
+    computeConnectorMethods() ;
+    computeRedondantRanks() ; 
+    for(auto& rank : rankToRemove_)
+      for(auto& element : elements_) element.erase(rank) ;
+  }
+/**
+  * \brief Compute the connector, i.e. compute the \b elements_ attribute. 
   * \detail In order to achive better optimisation,
   *         we distingute the case when the grid is not distributed on source grid (\bcomputeSrcNonDistributed), 
@@ -120 +133 @@
   *         redondantly the the remote servers
   */
-  void CGridRemoteConnector::computeConnector(void)
-  {
-    computeViewDistribution() ;
+  void CGridRemoteConnector::computeConnectorMethods(void)
+  {
     vector<CLocalView*> srcView ;
     vector<CDistributedView*> dstView ;
@@ -129 +141 @@
     
     bool srcViewsNonDistributed=true ;
-    for(int i=0;i<srcView_.size();i++) srcViewsNonDistributed &= !isSrcViewDistributed_[i]  ;
+    for(int i=0;i<srcView_.size();i++) srcViewsNonDistributed = srcViewsNonDistributed && !isSrcViewDistributed_[i]  ;
     
     bool dstViewsNonDistributed=true ;
-    for(int i=0;i<dstView_.size();i++) dstViewsNonDistributed &= !isDstViewDistributed_[i] ;
+    for(int i=0;i<dstView_.size();i++) dstViewsNonDistributed = dstViewsNonDistributed && !isDstViewDistributed_[i] ;
     
     if (srcViewsNonDistributed) 
@@ -184 +196 @@
     }
 
-    removeRedondantRanks() ;
   }
 
@@ -570 +581 @@
 
  /**
-  * \brief Once the connector is computed (compute \b elements_), redondant data can be send to the server. 
+  * \brief Once the connector is computed (compute \b elements_), redondant data can be avoid to be sent to the server. 
+  *        This call compute the redondant rank and store them in \b rankToRemove_ attribute.
   *        The goal of this method is to make a hash of each block of indice that determine wich data to send to a 
   *        of a specific server rank using a hash method. So data to send to a rank is associated to a hash.
   *        After we compare hash between local rank and remove redondant data corresponding to the same hash.
   */
-  void CGridRemoteConnector::removeRedondantRanks(void)
+  void CGridRemoteConnector::computeRedondantRanks(void)
   {
     int commRank ;
@@ -635 +647 @@
       // only the process with the lowest rank get in charge of sendinf data to remote server
       for(int rank : ranks) if (commRank>rank) first=false ;
-      if (!first)
-      {
-        int rankToRemove=hashRank[hash] ;
-        for(auto& element : elements_) element.erase(rankToRemove) ;
-      }
-    }
-
-
-  }
+      if (!first) rankToRemove_.insert(hashRank[hash]) ;
+    }
+  }
+
 }

XIOS/dev/dev_ym/XIOS_COUPLING/src/distribution/grid_remote_connector.hpp

@@ -22 +22 @@
       void computeViewDistribution(void) ;
       void computeConnector(void) ;
+      void computeConnectorMethods(void) ;
       void computeGenericMethod(vector<CLocalView*>& srcView, vector<CDistributedView*>& dstView, vector<int>& indElements) ;
       void computeSrcDstNonDistributed(int i, map<int,bool>& ranks) ;
       void computeDstNonDistributed(int i, map<int,bool>& ranks) ;
       void computeSrcNonDistributed(int i) ;
-      void removeRedondantRanks(void) ;
+      void computeRedondantRanks(void) ;
       std::map<int, CArray<size_t,1>>& getDistributedGlobalIndex(int pos) { return elements_[pos] ;} 
-
-    private:
+      const vector<bool>& getIsSrcViewDistributed(void) { return isSrcViewDistributed_ ;}
+      const vector<bool>& getIsDstViewDistributed(void) { return isDstViewDistributed_ ;}
+      const set<int>& getRankToRemove(void) {return rankToRemove_;} 
+    
+    protected:
   
     /** 
@@ -76 +80 @@
        */
       vector<bool> isDstViewDistributed_ ;
+
+      /**
+      /* Redondant ranks of the \b elements_ are stored there by calling computeRedondantRanks(), to be removed latter or to be retrieve from elsewhere. 
+       */
+      set<int> rankToRemove_ ;
       
     

XIOS/dev/dev_ym/XIOS_COUPLING/src/node/grid.cpp

@@ -18 +18 @@
 #include "server.hpp"
 #include "distribution_type.hpp"
-#include "grid_remote_connector.hpp"
+#include "grid_client_server_remote_connector.hpp"
 #include "grid_elements.hpp"
 #include "grid_local_view.hpp"
@@ -1429 +1429 @@
     
     vector<CLocalView*> localViews ;
+    vector<CLocalView*> workflowView ;
     vector<CDistributedView*> remoteViews ;
 
@@ -1439 +1440 @@
          remoteViews.push_back(domain->getRemoteElement(client)->getView(CElementView::FULL)) ;
          localViews.push_back(domain->getLocalView(CElementView::FULL)) ;
+         workflowView.push_back(domain->getLocalView(CElementView::WORKFLOW)) ;
       }
       else if (elements[i].type==TYPE_AXIS)
@@ -1446 +1448 @@
         remoteViews.push_back(axis->getRemoteElement(client)->getView(CElementView::FULL)) ;
         localViews.push_back(axis->getLocalView(CElementView::FULL)) ;
+        workflowView.push_back(axis->getLocalView(CElementView::WORKFLOW)) ;
       }
       else if (elements[i].type==TYPE_SCALAR)
@@ -1453 +1456 @@
         remoteViews.push_back(scalar->getRemoteElement(client)->getView(CElementView::FULL)) ;
         localViews.push_back(scalar->getLocalView(CElementView::FULL)) ;
-      }
-    }
-    CGridRemoteConnector gridRemoteConnector(localViews, remoteViews, context->getIntraComm(), client->getRemoteSize()) ;
+        workflowView.push_back(scalar->getLocalView(CElementView::WORKFLOW)) ;
+      }
+    }
+    
+    // CGridClientServerRemoteConnector : workflowView is added to avoid spurious optimisation with only the fullview
+    CGridClientServerRemoteConnector gridRemoteConnector(localViews, workflowView, remoteViews, context->getIntraComm(), client->getRemoteSize()) ;
     gridRemoteConnector.computeConnector() ;
     

XIOS/dev/dev_ym/XIOS_COUPLING/src/test/generic_testcase.f90

@@ -204 +204 @@
     LOGICAL :: ok_other_field0D_W, ok_other_field2D_W, ok_other_field3D_W, ok_other_pressure_W, ok_other_field2D_sub_W, ok_other_field3D_sub_W,ok_other_field3D_recv_W, ok_other_field3D_send_W
     LOGICAL :: ok_other_field_XW, ok_other_field_YW, ok_other_field_XYW, ok_other_field_ZW, ok_other_field_XYZW, ok_other_field_XZW, ok_other_field_YZW
+
+    TYPE(xios_domain)  :: domain_handle
     
       !! XIOS Initialization (get the local communicator)
@@ -635 +637 @@
     CALL xios_close_context_definition()
     CALL xios_set_current_context(trim(model_id))
-
+!    CALL xios_get_handle("field3D::domain",domain_handle) 
     
     DO while ( cdate <= edate )