Add Scheduler TImers (#109)

* Add schedule timing counters * Updates to loadbalance benchmark * doAll counters
utcs-scea · May 30, 2024 · de5af5c · de5af5c
1 parent f7558a3
commit de5af5c
Show file tree

Hide file tree

Showing 3 changed files with 36 additions and 11 deletions.
diff --git a/examples/benchmark_loadbalance.cpp b/examples/benchmark_loadbalance.cpp
@@ -55,20 +55,25 @@ int pandoMain(int argc, char** argv) {
     if (numItemsPerChunk != 1) {
       printUsageExit(argv[0]);
     }
-    Arr<G<std::uint64_t>> arr;
+    Arr<std::uint64_t> arr;
     PANDO_CHECK(arr.initialize(galois::getNumThreads()));
-    G<std::uint64_t> hello;
-    pando::LocalStorageGuard<std::uint64_t> helloGuard(hello, 1);
-    *hello = 0;
-    for (Ref<G<std::uint64_t>> item : arr) {
-      item = hello;
-    }
-    galois::doAll(
-        arr, +[](G<std::uint64_t> hello) {
-          pando::atomicFetchAdd(hello, 1, std::memory_order_relaxed);
-          while (pando::atomicLoad(hello, std::memory_order_relaxed) != galois::getNumThreads()) {}
+    galois::WaitGroup wg;
+    PANDO_CHECK(wg.initialize(0));
+    auto wgh = wg.getHandle();
+    counter::HighResolutionCount<true> makeSpanCounter;
+    makeSpanCounter.start();
+    galois::doAllEvenlyPartition(
+        wgh, arr, arr.size(),
+        +[](Arr<std::uint64_t> arr, std::uint64_t i, std::uint64_t numThreads) {
+          UNUSED(arr);
+          UNUSED(i);
+          UNUSED(numThreads);
         });
+    PANDO_CHECK(wg.wait());
+    auto time = makeSpanCounter.stop();
     arr.deinitialize();
+    wg.deinitialize();
+    std::cout << "The makespan of one task per thread was " << time.count() << std::endl;
   }
   pando::waitAll();
   return 0;

diff --git a/include/pando-lib-galois/loops/do_all.hpp b/include/pando-lib-galois/loops/do_all.hpp
@@ -18,10 +18,12 @@
 #include <pando-rt/pando-rt.hpp>
 
 constexpr bool SCHEDULER_TIMER_ENABLE = false;
+constexpr bool DOALL_TIMER_ENABLE = false;
 
 extern counter::Record<std::minstd_rand> perCoreRNG;
 extern counter::Record<std::uniform_int_distribution<std::int8_t>> perCoreDist;
 extern counter::Record<std::int64_t> schedulerCount;
+extern counter::Record<std::int64_t> doAllCount;
 
 namespace galois {
 
@@ -176,6 +178,8 @@ class DoAll {
   template <typename State, typename R, typename F, typename L>
   static pando::Status doAll(WaitGroup::HandleType wgh, State s, R& range, const F& func,
                              const L& localityFunc) {
+    counter::HighResolutionCount<DOALL_TIMER_ENABLE> doAllTimer;
+    doAllTimer.start();
     LoopLocalSchedulerStruct<CURRENT_SCHEDULER_POLICY> loopLocal{};
     pando::Status err = pando::Status::Success;
 
@@ -198,6 +202,7 @@ class DoAll {
       }
       iter++;
     }
+    counter::recordHighResolutionEvent(doAllCount, doAllTimer);
     return err;
   }
 
@@ -216,6 +221,8 @@ class DoAll {
    */
   template <typename State, typename R, typename F>
   static pando::Status doAll(WaitGroup::HandleType wgh, State s, R& range, const F& func) {
+    counter::HighResolutionCount<DOALL_TIMER_ENABLE> doAllTimer;
+    doAllTimer.start();
     LoopLocalSchedulerStruct<CURRENT_SCHEDULER_POLICY> loopLocal{};
     pando::Status err = pando::Status::Success;
 
@@ -238,6 +245,7 @@ class DoAll {
       }
       iter++;
     }
+    counter::recordHighResolutionEvent(doAllCount, doAllTimer);
     return err;
   }
 
@@ -256,6 +264,8 @@ class DoAll {
    */
   template <typename R, typename F>
   static pando::Status doAll(WaitGroup::HandleType wgh, R& range, const F& func) {
+    counter::HighResolutionCount<DOALL_TIMER_ENABLE> doAllTimer;
+    doAllTimer.start();
     LoopLocalSchedulerStruct<CURRENT_SCHEDULER_POLICY> loopLocal{};
     pando::Status err = pando::Status::Success;
 
@@ -277,6 +287,7 @@ class DoAll {
       }
       iter++;
     }
+    counter::recordHighResolutionEvent(doAllCount, doAllTimer);
     return err;
   }
 
@@ -370,9 +381,12 @@ class DoAll {
   template <typename State, typename F>
   static pando::Status doAllEvenlyPartition(WaitGroup::HandleType wgh, State s, uint64_t workItems,
                                             const F& func) {
+    counter::HighResolutionCount<DOALL_TIMER_ENABLE> doAllTimer;
+    doAllTimer.start();
     LoopLocalSchedulerStruct<EVENLY_PARITION_SCHEDULER_POLICY> loopLocal;
     pando::Status err = pando::Status::Success;
     if (workItems == 0) {
+      counter::recordHighResolutionEvent(doAllCount, doAllTimer);
       return err;
     }
 
@@ -407,6 +421,7 @@ class DoAll {
         return err;
       }
     }
+    counter::recordHighResolutionEvent(doAllCount, doAllTimer);
     return err;
   }
 

diff --git a/pando-rt/src/init.cpp b/pando-rt/src/init.cpp
@@ -35,6 +35,7 @@
 counter::Record<std::minstd_rand> perCoreRNG;
 counter::Record<std::uniform_int_distribution<std::int8_t>> perCoreDist;
 counter::Record<std::int64_t> schedulerCount = counter::Record<std::int64_t>();
+counter::Record<std::int64_t> doAllCount = counter::Record<std::int64_t>();
 
 namespace pando {
 
@@ -216,6 +217,10 @@ int main(int argc, char* argv[]) {
         thisPlace.node.id,
         std::int8_t((i == std::uint64_t(dims.core.x + 1)) ? -1 : i),
         schedulerCount.get(i));
+    SPDLOG_WARN("DoAll time on node: {}, core: {} was {}",
+        thisPlace.node.id,
+        std::int8_t((i == std::uint64_t(dims.core.x + 1)) ? -1 : i),
+        doAllCount.get(i));
   }