chore: introduce Idle measurement function (#323)

util/fibers/epoll_proactor.cc

@@ -264,7 +264,10 @@ void EpollProactor::MainLoop(detail::Scheduler* scheduler) {
       }
     }
 
+    uint64_t start_cycle = GetCPUCycleCount();
     int epoll_res = EpollWait(epoll_fd_, &ev_batch, timeout);
+    IdleEnd(start_cycle);
+
     if (epoll_res < 0) {
       epoll_res = errno;
       if (epoll_res == EINTR)

util/fibers/proactor_base.cc

@@ -54,10 +54,6 @@ void SigAction(int signal, siginfo_t*, void*) {
   }
 }
 
-inline uint64_t GetCPUCycleCount() {
-  return absl::base_internal::CycleClock::Now();
-}
-
 unsigned pause_amplifier = 50;
 uint64_t cycles_per_10us = 1000000;  // correctly defined inside ModuleInit.
 std::once_flag module_init;
@@ -293,6 +289,21 @@ bool ProactorBase::RunL2Tasks(detail::Scheduler* scheduler) {
   return result;
 }
 
+void ProactorBase::IdleEnd(uint64_t start) {
+  uint64_t end = GetCPUCycleCount();
+
+  // Assuming that cpu clock frequency is
+  uint64_t kMinCyclePeriod = cycles_per_10us * 500'000ULL;
+  cpu_idle_cycles_ += (end - start);
+
+  if (end > cpu_measure_cycle_start_ + kMinCyclePeriod) {
+    load_numerator_ = cpu_idle_cycles_;
+    load_denominator_ = end - cpu_measure_cycle_start_;
+    cpu_idle_cycles_ = 0;
+    cpu_measure_cycle_start_ = end;
+  }
+}
+
 void ProactorBase::Pause(unsigned count) {
   auto pc = pause_amplifier;
 
@@ -328,6 +339,7 @@ void ProactorBase::ModuleInit() {
 }
 
 void ProactorDispatcher::Run(detail::Scheduler* sched) {
+  proactor_->cpu_measure_cycle_start_ = ProactorBase::GetCPUCycleCount();
   proactor_->MainLoop(sched);
 }
 

util/fibers/proactor_base.h

@@ -21,6 +21,10 @@
 #include "util/fibers/fibers.h"
 #include "util/fibers/synchronization.h"
 
+#if !defined(__x86_64__) && !defined(__aarch64__)
+#include <absl/base/internal/cycleclock.h>
+#endif
+
 namespace util {
 class LinuxSocketBase;
 
@@ -204,6 +208,11 @@ class ProactorBase {
     return stats_;
   }
 
+  // Returns the idle ratio of the proactor thread. The number is between 0 and 1.
+  double IdleRatio() const {
+    return double(load_numerator_) / load_denominator_;
+  }
+
  protected:
   enum { WAIT_SECTION_STATE = 1UL << 31 };
   static constexpr unsigned kMaxSpinLimit = 5;
@@ -244,6 +253,7 @@ class ProactorBase {
     return absl::GetCurrentTimeNanos();
   }
 
+
   // Returns true if we should poll scheduler tasks that run periodically but not too often.
   bool ShouldPollL2Tasks() const;
 
@@ -252,6 +262,22 @@ class ProactorBase {
   // Returns true if there are fibers that became ready as a result.
   bool RunL2Tasks(detail::Scheduler* scheduler);
 
+  static uint64_t GetCPUCycleCount() {
+#if defined(__x86_64__)
+    uint64_t low, high;
+  __asm__ volatile("rdtsc" : "=a"(low), "=d"(high));
+  return static_cast<int64_t>((high << 32) | low);
+#elif defined(__aarch64__)
+  int64_t tv;
+  asm volatile("mrs %0, cntvct_el0" : "=r"(tv));
+  return tv;
+#else
+    return absl::base_internal::CycleClock::Now();
+#endif
+  }
+
+  void IdleEnd(uint64_t start);
+
   pthread_t thread_id_ = 0U;
   int sys_thread_id_ = 0;
   int32_t pool_index_ = -1;
@@ -310,6 +336,9 @@ class ProactorBase {
   }
 
   uint64_t last_level2_cycle_ = 0;
+
+  uint64_t cpu_measure_cycle_start_ = 0, cpu_idle_cycles_ = 0;
+  uint64_t load_numerator_ = 0, load_denominator_ = 1;
 };
 
 class ProactorDispatcher : public DispatchPolicy {

util/fibers/uring_proactor.cc

@@ -825,8 +825,9 @@ void UringProactor::MainLoop(detail::Scheduler* scheduler) {
       DCHECK(!scheduler->HasReady());
 
       VPRO(2) << "wait_for_cqe " << stats_.loop_cnt;
-
+      uint64_t start_cycle = GetCPUCycleCount();
       wait_for_cqe(&ring_, 1, ts_arg);
+      IdleEnd(start_cycle);
       VPRO(2) << "Woke up after wait_for_cqe ";
 
       ++stats_.num_stalls;