CPU util. measurements fix (more accurate and steady)

see #trex-246

5 files changed, 33 insertions, 16 deletions

diff --git a/src/bp_sim.cpp b/src/bp_sim.cpp
index 2d1d020b..773e82fc 100755
--- a/src/bp_sim.cpp
+++ b/src/bp_sim.cpp
@@ -4615,8 +4615,7 @@ double CFlowGenList::GetCpuUtilRaw(){
 
 void CFlowGenList::UpdateFast(){
 
-    int i;
-    for (i=0; i<(int)m_threads_info.size(); i++) {
+    for (int i=0; i<(int)m_threads_info.size(); i++) {
         CFlowGenListPerThread * lp=m_threads_info[i];
         lp->Update();
     }
diff --git a/src/main_dpdk.cpp b/src/main_dpdk.cpp
index 6beb4e3e..0ef83c02 100644
--- a/src/main_dpdk.cpp
+++ b/src/main_dpdk.cpp
@@ -4126,15 +4126,21 @@ void
 CGlobalTRex::handle_fast_path() {
     /* check from messages from DP */
     check_for_dp_messages();
-    // update CPU%
-    m_fl.UpdateFast();
 
-    if (get_is_stateless()) {
-        m_rx_sl.update_cpu_util();
-    }else{
-        m_mg.update_fast();
+    /* measure CPU utilization by sampling (we sample 1000 to get an accurate sampling) */
+    for (int i = 0; i < 1000; i++) {
+        m_fl.UpdateFast();
+
+        if (get_is_stateless()) {
+            m_rx_sl.update_cpu_util();
+        }else{
+            m_mg.update_fast();
+        }
+
+        rte_pause();
     }
 
+
     if ( is_all_cores_finished() ) {
         mark_for_shutdown(SHUTDOWN_TEST_ENDED);
     }
diff --git a/src/os_time.h b/src/os_time.h
index b1d44c03..4f10e2b5 100755
--- a/src/os_time.h
+++ b/src/os_time.h
@@ -165,6 +165,17 @@ void delay(int msec){
     nanosleep(&time1,&remain);
 }
 
-
+/**
+ * more accurate sleep by doing spin
+ * 
+ */
+static inline
+void delay_spin(double sec) {
+    double target = now_sec() + sec;
+
+    while (now_sec() < target) {
+        rte_pause();
+    }
+}
 
 #endif
diff --git a/src/utl_cpuu.cpp b/src/utl_cpuu.cpp
index 47c78c8e..c01326d6 100755
--- a/src/utl_cpuu.cpp
+++ b/src/utl_cpuu.cpp
@@ -42,10 +42,11 @@ void CCpuUtlCp::Update(){
     if ( m_dpcpu->sample_data() ) {
         m_work++;
     }
-    if (m_ticks==100) {
+    if (m_ticks==100000) {
         /* LPF*/
-        m_cpu_util_lpf = (m_cpu_util_lpf*0.75)+((double)m_work*0.25);
-        AppendHistory(m_work);
+        double work = (m_work / double(m_ticks)) * 100;
+        m_cpu_util_lpf = (m_cpu_util_lpf*0.75)+(work*0.25);
+        AppendHistory(work);
         m_ticks=0;
         m_work=0;
     }
diff --git a/src/utl_cpuu.h b/src/utl_cpuu.h
index b0a76fce..690eb200 100755
--- a/src/utl_cpuu.h
+++ b/src/utl_cpuu.h
@@ -47,7 +47,7 @@ public:
 
 
 private:
-    uint8_t m_data;
+    volatile uint8_t m_data;
 } __rte_cache_aligned; 
 
 class CCpuUtlCp {
@@ -62,9 +62,9 @@ public:
     void GetHistory(cpu_vct_st &cpu_vct);
 private:
     void AppendHistory(uint8_t);
-    CCpuUtlDp *         m_dpcpu;
-    uint8_t             m_ticks;
-    uint8_t             m_work;
+    CCpuUtlDp *          m_dpcpu;
+    uint32_t             m_ticks;
+    uint32_t             m_work;
 
     static const int    m_history_size=20;
     uint8_t             m_cpu_util[m_history_size]; // history as cyclic array