#708: Carry subphase timings into ProcStats

DARMA-tasking · Jun 8, 2020 · 2b78eec · 2b78eec
1 parent bfbbac9
commit 2b78eec
Show file tree

Hide file tree

Showing 3 changed files with 22 additions and 7 deletions.
diff --git a/src/vt/vrt/collection/balance/elm_stats.impl.h b/src/vt/vrt/collection/balance/elm_stats.impl.h
@@ -89,11 +89,12 @@ template <typename ColT>
   auto const& proxy = col->getCollectionProxy();
   auto const& untyped_proxy = col->getProxy();
   auto const& total_load = stats.getLoad(cur_phase, getFocusedSubPhase(untyped_proxy));
+  auto const& subphase_loads = stats.subphase_timings_.at(cur_phase);
   auto const& comm = stats.getComm(cur_phase);
   auto const& idx = col->getIndex();
   auto const& elm_proxy = proxy[idx];
 
-  theProcStats()->addProcStats<ColT>(elm_proxy, col, cur_phase, total_load, comm);
+  theProcStats()->addProcStats<ColT>(elm_proxy, col, cur_phase, total_load, subphase_loads, comm);
 
   auto const before_ready = theCollection()->numReadyCollections();
   theCollection()->makeCollectionReady(untyped_proxy);

diff --git a/src/vt/vrt/collection/balance/proc_stats.h b/src/vt/vrt/collection/balance/proc_stats.h
@@ -61,8 +61,9 @@
 namespace vt { namespace vrt { namespace collection { namespace balance {
 
 struct ProcStats : runtime::component::Component<ProcStats> {
-  using MigrateFnType = std::function<void(NodeType)>;
-  using LoadMapType   = std::unordered_map<ElementIDType,TimeType>;
+  using MigrateFnType       = std::function<void(NodeType)>;
+  using LoadMapType         = std::unordered_map<ElementIDType,TimeType>;
+  using SubphaseLoadMapType = std::unordered_map<ElementIDType, std::vector<TimeType>>;
 
   ProcStats() = default;
 
@@ -98,7 +99,8 @@ struct ProcStats : runtime::component::Component<ProcStats> {
   template <typename ColT>
   ElementIDType addProcStats(
     VirtualElmProxyType<ColT> const& elm_proxy, ColT* col_elm,
-    PhaseType const& phase, TimeType const& time, CommMapType const& comm
+    PhaseType const& phase, TimeType const& time,
+    std::vector<TimeType> const& subphase_time, CommMapType const& comm
   );
 
   /**
@@ -198,6 +200,8 @@ struct ProcStats : runtime::component::Component<ProcStats> {
   objgroup::proxy::Proxy<ProcStats> proxy_;
   /// Processor timings for each local object
   std::vector<LoadMapType> proc_data_;
+  /// Processor subphase timings for each local object
+  std::vector<SubphaseLoadMapType> proc_subphase_data_;
   /// Local migration type-free lambdas for each object
   std::unordered_map<ElementIDType,MigrateFnType> proc_migrate_;
   /// Map of temporary ID to permanent ID

diff --git a/src/vt/vrt/collection/balance/proc_stats.impl.h b/src/vt/vrt/collection/balance/proc_stats.impl.h
@@ -59,7 +59,8 @@ namespace vt { namespace vrt { namespace collection { namespace balance {
 template <typename ColT>
 ElementIDType ProcStats::addProcStats(
   VirtualElmProxyType<ColT> const& elm_proxy, ColT* col_elm,
-  PhaseType const& phase, TimeType const& time, CommMapType const& comm
+  PhaseType const& phase, TimeType const& time,
+  std::vector<TimeType> const& subphase_time, CommMapType const& comm
 ) {
   // A new temp ID gets assigned when a object is migrated into a node
 
@@ -68,8 +69,8 @@ ElementIDType ProcStats::addProcStats(
 
   debug_print(
     lb, node,
-    "ProcStats::addProcStats: temp_id={}, perm_id={}, phase={}, load={}\n",
-    temp_id, perm_id, phase, time
+    "ProcStats::addProcStats: temp_id={}, perm_id={}, phase={}, subphases={}, load={}\n",
+    temp_id, perm_id, phase, subphase_time.size(), time
   );
 
   proc_data_.resize(phase + 1);
@@ -81,6 +82,15 @@ ElementIDType ProcStats::addProcStats(
     std::forward_as_tuple(time)
   );
 
+  proc_subphase_data_.resize(phase + 1);
+  auto elm_subphase_iter = proc_subphase_data_.at(phase).find(temp_id);
+  vtAssert(elm_subphase_iter == proc_subphase_data_.at(phase).end(), "Must not exist");
+  proc_subphase_data_.at(phase).emplace(
+    std::piecewise_construct,
+    std::forward_as_tuple(temp_id),
+    std::forward_as_tuple(subphase_time)
+  );
+
   proc_comm_.resize(phase + 1);
   for (auto&& c : comm) {
     proc_comm_.at(phase)[c.first] += c.second;