Merge pull request #2333 from DARMA-tasking/2299-dont-deploy-lb-with-…

…to-small-average-load

#2299: Do not deploy LB when average load is smaller than estimated load balancing cost

src/vt/vrt/collection/balance/baselb/baselb.h

@@ -55,6 +55,7 @@
 #include <map>
 #include <unordered_map>
 #include <tuple>
+#include <chrono>
 
 namespace vt { namespace vrt { namespace collection {
 
@@ -188,6 +189,32 @@ struct BaseLB {
     std::shared_ptr<balance::LoadModel> model
   );
 
+  /**
+   * \brief Get the estimated time needed for load-balancing
+   *
+   * \return the estimated time
+   */
+  auto getCollectiveEpochCost() const {
+    return std::chrono::nanoseconds(100);
+  }
+
+  /**
+   * \brief Check if load-balancing should be done
+   *
+   * \return true when the maximum load exceeds the cost of load balancing; false otherwise
+   */
+  bool maxLoadExceedsLBCost() const {
+    auto const max = base_stats_->at(lb::Statistic::Rank_load_modeled).at(
+      lb::StatisticQuantity::max
+    );
+    auto max_in_ns = std::chrono::duration_cast<std::chrono::nanoseconds>(
+      std::chrono::duration<double>(max)
+    );
+
+    // Compare the maximum rank load to the estimated load-balancing cost
+    return max_in_ns > getCollectiveEpochCost();
+  }
+
 private:
   TransferVecType transfers_                      = {};
   TransferType off_node_migrate_                  = {};

src/vt/vrt/collection/balance/greedylb/greedylb.cc

@@ -160,7 +160,7 @@ void GreedyLB::loadStats() {
   bool should_lb = false;
   this_load_begin = this_load;
 
-  if (avg_load > 0.0000000001) {
+  if (maxLoadExceedsLBCost()) {
     should_lb = I > greedy_tolerance;
   }
 

src/vt/vrt/collection/balance/hierarchicallb/hierlb.cc

@@ -263,7 +263,7 @@ void HierarchicalLB::loadStats() {
   bool should_lb = false;
   this_load_begin = this_load;
 
-  if (avg_load > 0.0000000001) {
+  if (maxLoadExceedsLBCost()) {
     should_lb = I > hierlb_tolerance;
   }
 

src/vt/vrt/collection/balance/temperedlb/temperedlb.cc

@@ -449,7 +449,7 @@ void TemperedLB::runLB(LoadType total_load) {
     target_max_load_ = avg;
   }
 
-  if (avg > 0.0000000001) {
+  if (maxLoadExceedsLBCost()) {
     should_lb = max > (run_temperedlb_tolerance + 1.0) * target_max_load_;
   }