Adapted perfomance counters for AMD Epyc and usage of TSC instead of std::chrono.

src/application/blinktree_benchmark/benchmark.cpp

@@ -23,9 +23,15 @@ Benchmark::Benchmark(benchmark::Cores &&cores, const std::uint16_t iterations, s
     {
         this->_chronometer.add(benchmark::Perf::CYCLES);
         this->_chronometer.add(benchmark::Perf::INSTRUCTIONS);
-        this->_chronometer.add(benchmark::Perf::STALLS_MEM_ANY);
+        //this->_chronometer.add(benchmark::Perf::STALLS_MEM_ANY);
-        this->_chronometer.add(benchmark::Perf::SW_PREFETCH_ACCESS_NTA);
+        //this->_chronometer.add(benchmark::Perf::SW_PREFETCH_ACCESS_NTA);
-        this->_chronometer.add(benchmark::Perf::SW_PREFETCH_ACCESS_WRITE);
+        //this->_chronometer.add(benchmark::Perf::SW_PREFETCH_ACCESS_WRITE);
+        this->_chronometer.add(benchmark::Perf::LLC_MISSES);
+        this->_chronometer.add(benchmark::Perf::DTLB_READ_MISSES);
+        this->_chronometer.add(benchmark::Perf::DTLB_STORE_MISSES);
+        this->_chronometer.add(benchmark::Perf::ITLB_LOAD_MISSES);
+        this->_chronometer.add(benchmark::Perf::SW_PAGE_FAULTS_MAJOR);
+        this->_chronometer.add(benchmark::Perf::SW_PAGE_FAULTS_MINOR);
     }
 
     std::cout << "core configuration: \n" << this->_cores.dump(2) << std::endl;
@@ -53,7 +59,10 @@ void Benchmark::start()
     {
         this->_request_scheduler.clear();
     }
-
+    
+    auto *start_task = mx::tasking::runtime::new_task<StartMeasurementTask>(0U, *this);
+    mx::tasking::runtime::spawn(*start_task, 0U);
+    
     // Create one request scheduler per core.
     for (auto core_index = 0U; core_index < this->_cores.current().size(); core_index++)
     {
@@ -70,8 +79,8 @@ void Benchmark::start()
     {
         mx::tasking::runtime::profile(this->profile_file_name());
     }
-    this->_chronometer.start(static_cast<std::uint16_t>(static_cast<benchmark::phase>(this->_workload)),
+    //this->_chronometer.start(static_cast<std::uint16_t>(static_cast<benchmark::phase>(this->_workload)),
-                             this->_current_iteration + 1, this->_cores.current());
+      //                       this->_current_iteration + 1, this->_cores.current());
 }
 
 const mx::util::core_set &Benchmark::core_set()
@@ -109,6 +118,14 @@ void Benchmark::requests_finished()
 
     if (open_requests == 0U) // All request schedulers are done.
     {
+        std::uint16_t core_id = mx::system::topology::core_id();
+        if (core_id != 0) { 
+            this->_open_requests++;
+            auto *stop_task = mx::tasking::runtime::new_task<StopMeasurementTask>(0U, *this);
+            stop_task->annotate(static_cast<mx::tasking::TaskInterface::channel>(0));
+            mx::tasking::runtime::spawn(*stop_task, core_id);
+            return;
+        }
         // Stop and print time (and performance counter).
         const auto result = this->_chronometer.stop(this->_workload.size());
         mx::tasking::runtime::stop();
@@ -193,7 +210,7 @@ void Benchmark::requests_finished()
 
 std::string Benchmark::profile_file_name() const
 {
-    return "profiling-" + std::to_string(this->_cores.current().size()) + "-cores" + "-phase-" +
+    return "profiling-" + std::to_string(static_cast<int>(this->_cores.current().size())) + "-cores" + "-phase-" +
-           std::to_string(static_cast<std::uint16_t>(static_cast<benchmark::phase>(this->_workload))) + "-iteration-" +
+           std::to_string(static_cast<int>(static_cast<benchmark::phase>(this->_workload))) + "-iteration-" +
-           std::to_string(this->_current_iteration) + ".json";
+           std::to_string(static_cast<int>(this->_current_iteration)) + ".json";
-}
+}

src/benchmark/chronometer.h

@@ -7,6 +7,7 @@
 #include <mx/tasking/config.h>
 #include <mx/tasking/profiling/statistic.h>
 #include <mx/tasking/runtime.h>
+#include <mx/system/environment.h>
 #include <mx/util/core_set.h>
 #include <numeric>
 #include <ostream>
@@ -177,15 +178,16 @@ public:
         _core_set = core_set;
 
         _perf.start();
-        _start = std::chrono::steady_clock::now();
+        _start = mx::system::Environment::timestamp(); // std::chrono::steady_clock::now();
     }
 
     InterimResult<P> stop(const std::uint64_t count_operations)
     {
-        const auto end = std::chrono::steady_clock::now();
+        const auto end = mx::system::Environment::timestamp();  // std::chrono::steady_clock::now();
         _perf.stop();
 
-        const auto milliseconds = std::chrono::duration_cast<std::chrono::milliseconds>(end - _start);
+        const auto milliseconds = std::chrono::milliseconds(
+            (end - _start) / 2000000UL); // std::chrono::duration_cast<std::chrono::milliseconds>(end - _start);
 
         return {count_operations,
                 _current_phase,
@@ -209,7 +211,8 @@ private:
     P _current_phase;
     mx::util::core_set _core_set;
     alignas(64) Perf _perf;
-    alignas(64) std::chrono::steady_clock::time_point _start;
+    //alignas(64) std::chrono::steady_clock::time_point _start;
+    alignas(64) uint64_t _start;
 
     std::unordered_map<std::uint16_t, std::uint64_t> statistic_map(
         const mx::tasking::profiling::Statistic::Counter counter)

src/benchmark/perf.cpp

@@ -24,6 +24,13 @@ using namespace benchmark;
  */
 [[maybe_unused]] PerfCounter Perf::LLC_MISSES = {"llc-miss", PERF_TYPE_HARDWARE, PERF_COUNT_HW_CACHE_MISSES};
 
+[[maybe_unused]] PerfCounter Perf::DTLB_READ_MISSES = {"dtlb-read-miss", PERF_TYPE_HW_CACHE, 0x10003};
+[[maybe_unused]] PerfCounter Perf::DTLB_STORE_MISSES = {"dtlb-store-miss", PERF_TYPE_HW_CACHE, 0x10103};
+[[maybe_unused]] PerfCounter Perf::ITLB_LOAD_MISSES = {"itlb-load-miss", PERF_TYPE_HW_CACHE, 0x10004};
+[[maybe_unused]] PerfCounter Perf::SW_PAGE_FAULTS = {"sw-page-faults", PERF_TYPE_SOFTWARE, PERF_COUNT_SW_PAGE_FAULTS};
+[[maybe_unused]] PerfCounter Perf::SW_PAGE_FAULTS_MINOR = {"sw-page-faults-minor", PERF_TYPE_SOFTWARE, PERF_COUNT_SW_PAGE_FAULTS_MIN};
+[[maybe_unused]] PerfCounter Perf::SW_PAGE_FAULTS_MAJOR = {"sw-page-faults-major", PERF_TYPE_SOFTWARE, PERF_COUNT_SW_PAGE_FAULTS_MAJ};
+
 /**
  * Counter "LLC Reference"
  * Accesses to the LLC, in which the data is present(hit) or not present(miss)

src/benchmark/perf.h

@@ -107,6 +107,12 @@ public:
     [[maybe_unused]] static PerfCounter SW_PREFETCH_ACCESS_T0;
     [[maybe_unused]] static PerfCounter SW_PREFETCH_ACCESS_T1_T2;
     [[maybe_unused]] static PerfCounter SW_PREFETCH_ACCESS_WRITE;
+    [[maybe_unused]] static PerfCounter DTLB_STORE_MISSES;
+    [[maybe_unused]] static PerfCounter DTLB_READ_MISSES;
+    [[maybe_unused]] static PerfCounter ITLB_LOAD_MISSES;
+    [[maybe_unused]] static PerfCounter SW_PAGE_FAULTS;
+    [[maybe_unused]] static PerfCounter SW_PAGE_FAULTS_MINOR;
+    [[maybe_unused]] static PerfCounter SW_PAGE_FAULTS_MAJOR;
 
     Perf() noexcept = default;
     ~Perf() noexcept = default;
@@ -154,4 +160,4 @@ public:
 private:
     std::vector<PerfCounter> _counter;
 };
-} // namespace benchmark
+} // namespace benchmark