AnyCore now randomly selects a core at the executeOn level. (#119)

* AnyCore now randomly selects a core at the executeOn level. * bugfix
utcs-scea · Jun 14, 2024 · 56256b2 · 56256b2
1 parent 6478593
commit 56256b2
Show file tree

Hide file tree

Showing 4 changed files with 26 additions and 19 deletions.
diff --git a/pando-rt/include/pando-rt/execution/execute_on_impl.hpp b/pando-rt/include/pando-rt/execution/execute_on_impl.hpp
@@ -8,6 +8,11 @@
 #include "../status.hpp"
 #include "export.h"
 #include "task.hpp"
+#include <pando-rt/benchmark/counters.hpp>
+#include <random>
+
+extern counter::Record<std::minstd_rand> perCoreRNG;
+extern counter::Record<std::uniform_int_distribution<std::int8_t>> perCoreDist;
 
 namespace pando {
 

diff --git a/pando-rt/src/execute_on.cpp b/pando-rt/src/execute_on.cpp
@@ -30,8 +30,10 @@ Status detail::executeOn(Place place, Task task) {
 
   const auto& coreDims = getCoreDims();
   if (place.core == anyCore) {
+    auto coreIdx = perCoreDist.getLocal()(perCoreRNG.getLocal());
+    assert(coreIdx < pando::getCoreDims().x);
     // anyCore: get scheduler core queue
-    const CoreIndex schedulerCoreIndex(coreDims.x, 0);
+    const CoreIndex schedulerCoreIndex(coreIdx, 0);
     place.core = schedulerCoreIndex;
   } else {
     if ((place.core < CoreIndex{0, 0}) ||

diff --git a/pando-rt/src/start.cpp b/pando-rt/src/start.cpp
@@ -22,7 +22,7 @@
 #include "drvx/drvx.hpp"
 #endif
 
-constexpr std::uint64_t STEAL_THRESH_HOLD_SIZE = 32;
+constexpr std::uint64_t STEAL_THRESH_HOLD_SIZE = 16;
 
 constexpr bool IDLE_TIMER_ENABLE = false;
 counter::Record<std::int64_t> idleCount = counter::Record<std::int64_t>();
@@ -80,7 +80,7 @@ extern "C" int __start(int argc, char** argv) {
               break;
 
             case SchedulerFailState::STEAL:
-              for(std::int8_t i = 0; i <= coreDims.x && !task.has_value(); i++) {
+              for(std::int8_t i = 0; i < coreDims.x && !task.has_value(); i++) {
                 auto* otherQueue =  pando::Cores::getTaskQueue(pando::Place{thisPlace.node, thisPlace.pod, pando::CoreIndex(i, 0)});
                 if(!otherQueue || otherQueue == queue) {continue;}
                 if(otherQueue->getApproxSize() > STEAL_THRESH_HOLD_SIZE) {

diff --git a/pando-rt/test/execution/test_bulk_execute_on.cpp b/pando-rt/test/execution/test_bulk_execute_on.cpp
@@ -15,25 +15,25 @@
 
 namespace {
 
-pando::NodeSpecificStorage<std::int64_t> counter;
+pando::NodeSpecificStorage<std::int64_t> nodeCounter;
 
 struct IncrementCounter {
   void operator()(std::int64_t val) {
-    pando::atomicIncrement(counter.getPointer(), val, std::memory_order_relaxed);
+    pando::atomicIncrement(nodeCounter.getPointer(), val, std::memory_order_relaxed);
   }
 };
 
 struct IncrementCounters {
   void operator()(std::int64_t x, std::int64_t y) {
-    pando::atomicIncrement(&counter, x, std::memory_order_relaxed);
-    pando::atomicIncrement(&counter, y, std::memory_order_relaxed);
+    pando::atomicIncrement(&nodeCounter, x, std::memory_order_relaxed);
+    pando::atomicIncrement(&nodeCounter, y, std::memory_order_relaxed);
   }
 };
 
 } // namespace
 
 TEST(BulkExecuteOn, ThisNode) {
-  counter = 0;
+  nodeCounter = 0;
 
   const auto& thisPlace = pando::getCurrentPlace();
   const pando::Place place{thisPlace.node, pando::PodIndex{0, 0}, pando::CoreIndex{0, 0}};
@@ -43,42 +43,42 @@ TEST(BulkExecuteOn, ThisNode) {
             pando::Status::Success);
 
   pando::waitUntil([&] {
-    return pando::atomicLoad(&counter, std::memory_order_relaxed) == 10;
+    return pando::atomicLoad(&nodeCounter, std::memory_order_relaxed) == 10;
   });
 
-  EXPECT_EQ(pando::atomicLoad(&counter, std::memory_order_relaxed), 10);
+  EXPECT_EQ(pando::atomicLoad(&nodeCounter, std::memory_order_relaxed), 10);
 }
 
 TEST(BulkExecuteOn, ThisNodeAnyPod) {
-  counter = 0;
+  nodeCounter = 0;
 
   EXPECT_EQ(pando::bulkExecuteOn(pando::anyPod, IncrementCounter{}, std::make_tuple(1),
                                  std::make_tuple(2), std::make_tuple(3), std::make_tuple(4)),
             pando::Status::Success);
 
   pando::waitUntil([&] {
-    return pando::atomicLoad(&counter, std::memory_order_relaxed) == 10;
+    return pando::atomicLoad(&nodeCounter, std::memory_order_relaxed) == 10;
   });
 
-  EXPECT_EQ(pando::atomicLoad(&counter, std::memory_order_relaxed), 10);
+  EXPECT_EQ(pando::atomicLoad(&nodeCounter, std::memory_order_relaxed), 10);
 }
 
 TEST(BulkExecuteOn, ThisNodeAnyCore) {
-  counter = 0;
+  nodeCounter = 0;
 
   EXPECT_EQ(pando::bulkExecuteOn(pando::anyCore, IncrementCounter{}, std::make_tuple(1),
                                  std::make_tuple(2), std::make_tuple(3), std::make_tuple(4)),
             pando::Status::Success);
 
   pando::waitUntil([&] {
-    return pando::atomicLoad(&counter, std::memory_order_relaxed) == 10;
+    return pando::atomicLoad(&nodeCounter, std::memory_order_relaxed) == 10;
   });
 
-  EXPECT_EQ(pando::atomicLoad(&counter, std::memory_order_relaxed), 10);
+  EXPECT_EQ(pando::atomicLoad(&nodeCounter, std::memory_order_relaxed), 10);
 }
 
 TEST(BulkExecuteOn, ThisNodeMultipleArgs) {
-  counter = 0;
+  nodeCounter = 0;
 
   const auto& thisPlace = pando::getCurrentPlace();
   const pando::Place place{thisPlace.node, pando::PodIndex{0, 0}, pando::CoreIndex{0, 0}};
@@ -89,8 +89,8 @@ TEST(BulkExecuteOn, ThisNodeMultipleArgs) {
       pando::Status::Success);
 
   pando::waitUntil([&] {
-    return pando::atomicLoad(&counter, std::memory_order_relaxed) == 20;
+    return pando::atomicLoad(&nodeCounter, std::memory_order_relaxed) == 20;
   });
 
-  EXPECT_EQ(pando::atomicLoad(&counter, std::memory_order_relaxed), 20);
+  EXPECT_EQ(pando::atomicLoad(&nodeCounter, std::memory_order_relaxed), 20);
 }