Account for CPU used by child processes

getCpuAndContexts() only sampled RUSAGE_SELF, so CPU burned by forked
children was never reported. Sample RUSAGE_CHILDREN as well.

The final measurement is reachable via finalizeProcessMonitoring()
instead of only ~Monitoring(), and bypasses the 1s rate guard, which
would otherwise discard a delta that no later call can pick up.

Related to https://its.cern.ch/jira/browse/O2-7096

Assisted by Claude Opus 5

Author: sawenzel

include/Monitoring/Monitoring.h

@@ -73,6 +73,10 @@ class Monitoring
   /// \param enabledMeasurements vector of monitor measurements, eg. PmMeasurement::Cpu
   void enableProcessMonitoring(const unsigned int interval = 5, std::vector<PmMeasurement> enabledMeasurements = {PmMeasurement::Cpu, PmMeasurement::Mem, PmMeasurement::Smaps});
 
+  /// Stops process monitoring and transmits the final measurement. Idempotent;
+  /// call explicitly where destructor timing is not guaranteed to be reached.
+  void finalizeProcessMonitoring();
+
   /// Flushes metric buffer (this can also happen when buffer is full)
   void flushBuffer();
 

include/Monitoring/ProcessMonitor.h

@@ -112,6 +112,9 @@ class ProcessMonitor
   /// Best-effort open of the retired-instructions counter (no-op off Linux)
   void openInstructionCounter();
 
+  /// 'getrusage(RUSAGE_CHILDREN)' values from last execution
+  struct rusage mPreviousGetrUsageChildren;
+
   ///each measurement will be saved to compute average/accumulation usage
   std::vector<double> mVmSizeMeasurements;
   std::vector<double> mVmRssMeasurements;
@@ -128,7 +131,9 @@ class ProcessMonitor
   std::vector<Metric> getSmaps();
 
   /// Retrieves CPU usage (%) and number of context switches during the interval
-  std::vector<Metric> getCpuAndContexts();
+  /// \param force  ignore the 1s minimum interval; for the final measurement,
+  ///               where a skipped delta would be lost rather than deferred
+  std::vector<Metric> getCpuAndContexts(bool force = false);
 
   std::vector<Metric> makeLastMeasurementAndGetMetrics();
 };

src/Monitoring.cxx

@@ -130,13 +130,21 @@ void Monitoring::addBackend(std::unique_ptr<Backend> backend)
   mBackends.push_back(std::move(backend));
 }
 
-Monitoring::~Monitoring()
+void Monitoring::finalizeProcessMonitoring()
 {
+  if (!mMonitorRunning) {
+    return;
+  }
   mMonitorRunning = false;
   if (mMonitorThread.joinable()) {
     mMonitorThread.join();
     transmit(mProcessMonitor->makeLastMeasurementAndGetMetrics());
   }
+}
+
+Monitoring::~Monitoring()
+{
+  finalizeProcessMonitoring();
   flushBuffer();
 }
 

src/ProcessMonitor.cxx

@@ -60,6 +60,7 @@ ProcessMonitor::ProcessMonitor()
   mPid = static_cast<unsigned int>(::getpid());
   mTimeLastRun = std::chrono::high_resolution_clock::now();
   getrusage(RUSAGE_SELF, &mPreviousGetrUsage);
+  getrusage(RUSAGE_CHILDREN, &mPreviousGetrUsageChildren);
 #ifdef O2_MONITORING_OS_LINUX
   setTotalMemory();
 #endif
@@ -99,6 +100,7 @@ void ProcessMonitor::init()
 {
   mTimeLastRun = std::chrono::high_resolution_clock::now();
   getrusage(RUSAGE_SELF, &mPreviousGetrUsage);
+  getrusage(RUSAGE_CHILDREN, &mPreviousGetrUsageChildren);
 }
 
 void ProcessMonitor::enable(PmMeasurement measurement)
@@ -167,27 +169,41 @@ std::vector<Metric> ProcessMonitor::getSmaps()
   return {{pssTotal, metricsNames[PSS]}, {cleanTotal, metricsNames[PRIVATE_CLEAN]}, {dirtyTotal, metricsNames[PRIVATE_DIRTY]}};
 }
 
-std::vector<Metric> ProcessMonitor::getCpuAndContexts()
+std::vector<Metric> ProcessMonitor::getCpuAndContexts(bool force)
 {
   std::vector<Metric> metrics;
   struct rusage currentUsage;
+  struct rusage currentUsageChildren;
   getrusage(RUSAGE_SELF, &currentUsage);
+  // CPU of reaped children (e.g. an external event generator forked by o2-sim)
+  // is spent outside this process and is invisible to RUSAGE_SELF
+  getrusage(RUSAGE_CHILDREN, &currentUsageChildren);
   auto timeNow = std::chrono::high_resolution_clock::now();
   double timePassed = std::chrono::duration_cast<std::chrono::microseconds>(timeNow - mTimeLastRun).count();
-  if (timePassed < 950) {
+  if (timePassed < 950 && !force) {
     MonLogger::Get(Severity::Warn) << "Do not invoke Process Monitor more frequent then every 1s" << MonLogger::End();
     metrics.emplace_back("processPerformance");
     return metrics;
   }
 
-  uint64_t cpuUsedInMicroSeconds = currentUsage.ru_utime.tv_sec * 1000000.0 + currentUsage.ru_utime.tv_usec - (mPreviousGetrUsage.ru_utime.tv_sec * 1000000.0 + mPreviousGetrUsage.ru_utime.tv_usec) + currentUsage.ru_stime.tv_sec * 1000000.0 + currentUsage.ru_stime.tv_usec - (mPreviousGetrUsage.ru_stime.tv_sec * 1000000.0 + mPreviousGetrUsage.ru_stime.tv_usec);
+  auto micros = [](const timeval& t) { return t.tv_sec * 1000000.0 + t.tv_usec; };
+  auto cpuDelta = [&micros](const struct rusage& now, const struct rusage& before) {
+    return micros(now.ru_utime) - micros(before.ru_utime) + micros(now.ru_stime) - micros(before.ru_stime);
+  };
+  uint64_t cpuUsedInMicroSeconds = cpuDelta(currentUsage, mPreviousGetrUsage) +
+                                   cpuDelta(currentUsageChildren, mPreviousGetrUsageChildren);
   double fractionCpuUsed = cpuUsedInMicroSeconds / timePassed;
 
   double cpuUsedPerctange = std::round(fractionCpuUsed * 100.0 * 100.0) / 100.0;
-  mCpuPerctange.push_back(cpuUsedPerctange);
   mCpuMicroSeconds.push_back(cpuUsedInMicroSeconds);
 
-  metrics.emplace_back(Metric{cpuUsedPerctange, metricsNames[CPU_USED_PERCENTAGE]});
+  // A forced measurement may report CPU accumulated over the whole run but only
+  // made visible at once (children become visible on reap), for which an
+  // instantaneous rate is meaningless: report it as absolute time only.
+  if (!force) {
+    mCpuPerctange.push_back(cpuUsedPerctange);
+    metrics.emplace_back(Metric{cpuUsedPerctange, metricsNames[CPU_USED_PERCENTAGE]});
+  }
   metrics.emplace_back(Metric{
     static_cast<uint64_t>(currentUsage.ru_nivcsw - mPreviousGetrUsage.ru_nivcsw), metricsNames[INVOLUNTARY_CONTEXT_SWITCHES]});
   metrics.emplace_back(Metric{
@@ -212,6 +228,7 @@ std::vector<Metric> ProcessMonitor::getCpuAndContexts()
 
   mTimeLastRun = timeNow;
   mPreviousGetrUsage = currentUsage;
+  mPreviousGetrUsageChildren = currentUsageChildren;
   return metrics;
 }
 
@@ -262,7 +279,9 @@ std::vector<Metric> ProcessMonitor::makeLastMeasurementAndGetMetrics()
   }
 #endif
   if (mEnabledMeasurements.at(static_cast<short>(PmMeasurement::Cpu))) {
-    getCpuAndContexts();
+    // forced: no later call will pick up a delta discarded here
+    auto lastCpuMetrics = getCpuAndContexts(true);
+    std::move(lastCpuMetrics.begin(), lastCpuMetrics.end(), std::back_inserter(metrics));
 
     auto avgCpuUsage = std::accumulate(mCpuPerctange.begin(), mCpuPerctange.end(), 0.0) /
                        mCpuPerctange.size();