#2240: Fix runtime failure in allreduce perf test

DARMA-tasking · Oct 10, 2024 · 51ea74e · 51ea74e
1 parent 74bea76
commit 51ea74e
Show file tree

Hide file tree

Showing 2 changed files with 10 additions and 15 deletions.
diff --git a/src/vt/collective/reduce/allreduce/rabenseifner.impl.h b/src/vt/collective/reduce/allreduce/rabenseifner.impl.h
@@ -132,6 +132,7 @@ template <
 >
 void Rabenseifner<DataT, Op, ObjT, finalHandler>::executeFinalHan() {
   // theCB()->makeSend<finalHandler>(parent_proxy_[this_node_]).sendTuple(std::make_tuple(val_));
+  vt_debug_print(terse, allreduce, "Rabenseifner executing final handler\n");
   parent_proxy_[this_node_].template invoke<finalHandler>(val_);
   completed_ = true;
 }
@@ -267,7 +268,7 @@ void Rabenseifner<DataT, Op, ObjT, finalHandler>::scatterReduceIter() {
   auto dest = (vdest < nprocs_rem_) ? vdest * 2 : vdest + nprocs_rem_;
   vt_debug_print(
     terse, allreduce,
-    "Rabenseifner Part2 (step {}): Sending to Node {} starting with idx = {} and "
+    "Rabenseifner Part2 (Send step {}): To Node {} starting with idx = {} and "
     "count "
     "{} \n",
     scatter_step_, dest, s_index_[scatter_step_],
@@ -310,7 +311,7 @@ void Rabenseifner<DataT, Op, ObjT, finalHandler>::scatterReduceIterHandler(
 
   vt_debug_print(
     terse, allreduce,
-    "Rabenseifner Part2 (step {}): scatter_mask_= {} nprocs_pof2_ = {}: "
+    "Rabenseifner Part2 (Recv step {}): scatter_mask_= {} nprocs_pof2_ = {}: "
     "idx = {} from {}\n",
     msg->step_, scatter_mask_, nprocs_pof2_, r_index_[msg->step_],
     theContext()->getFromNodeCurrentTask()

diff --git a/tests/perf/allreduce.cc b/tests/perf/allreduce.cc
@@ -58,8 +58,11 @@
 using namespace vt;
 using namespace vt::tests::perf::common;
 
-static constexpr std::array<size_t, 8> const payloadSizes = {
-  64, 128, 2048, 16384, 32768, 524288, 1048576, 2097152};
+// static constexpr std::array<size_t, 8> const payloadSizes = {
+//   64, 128, 2048, 16384, 32768, 524288, 1048576, 2097152};
+
+static constexpr std::array<size_t, 1> const payloadSizes = {
+2097152};
 
 struct MyTest : PerfTestHarness {
   MyTest() {
@@ -109,7 +112,6 @@ VT_PERF_TEST(MyTest, test_reduce) {
     data.resize(payload_size, theContext()->getNode() + 1);
 
     theCollective()->barrier();
-
     StartTimer(grp_proxy[my_node_].get()->timer_names_.at(payload_size));
     grp_proxy.allreduce<&NodeObj::handlerVec, collective::PlusOp>(data);
   }
@@ -124,16 +126,12 @@ VT_PERF_TEST(MyTest, test_allreduce_rabenseifner) {
   using Reducer = collective::reduce::allreduce::Rabenseifner<
     DataT, collective::PlusOp, NodeObj, &NodeObj::handlerVec>;
 
-  auto grp_proxy = vt::theObjGroup()->makeCollective<Reducer>(
-    "allreduce_rabenseifner", proxy, num_nodes_, data);
-  grp_proxy[my_node_].get()->proxy_ = grp_proxy;
-
   for (auto payload_size : payloadSizes) {
     data.resize(payload_size, theContext()->getNode() + 1);
 
     theCollective()->barrier();
     StartTimer(proxy[my_node_].get()->timer_names_.at(payload_size));
-    grp_proxy[my_node_].template invoke<&Reducer::allreduce>();
+    proxy.allreduce_h<&NodeObj::handlerVec, collective::PlusOp>(data);
   }
 }
 
@@ -146,16 +144,12 @@ VT_PERF_TEST(MyTest, test_allreduce_recursive_doubling) {
   using Reducer = collective::reduce::allreduce::RecursiveDoubling<
     DataT, collective::PlusOp, NodeObj, &NodeObj::handlerVec>;
 
-  auto grp_proxy = vt::theObjGroup()->makeCollective<Reducer>(
-    "allreduce_recursive_doubling", proxy, num_nodes_, data);
-  grp_proxy[my_node_].get()->proxy_ = grp_proxy;
-
   for (auto payload_size : payloadSizes) {
     data.resize(payload_size, theContext()->getNode() + 1);
 
     theCollective()->barrier();
     StartTimer(proxy[my_node_].get()->timer_names_.at(payload_size));
-    grp_proxy[my_node_].template invoke<&Reducer::allreduce>();
+    proxy.allreduce_h<&NodeObj::handlerVec, collective::PlusOp>(data);
   }
 }