cpu% without tsc

9 files changed, 55 insertions, 49 deletions

diff --git a/src/bp_sim.cpp b/src/bp_sim.cpp
index f9cb3220..ccd62f0e 100755
--- a/src/bp_sim.cpp
+++ b/src/bp_sim.cpp
@@ -3727,6 +3727,9 @@ inline int CNodeGenerator::teardown(CFlowGenListPerThread * thread,
                                            bool always,
                                            double &old_offset,
                                            double offset){
+
+    thread->m_cpu_dp_u.commit1();
+
         /* to do */
     if ( thread->is_terminated_by_master() ) {
         return (0);
@@ -4511,7 +4514,7 @@ double CFlowGenList::GetCpuUtil(){
 }
 
 
-void CFlowGenList::Update(){
+void CFlowGenList::UpdateFast(){
 
     int i;
     for (i=0; i<(int)m_threads_info.size(); i++) {
diff --git a/src/bp_sim.h b/src/bp_sim.h
index bdca7703..6c8911e5 100755
--- a/src/bp_sim.h
+++ b/src/bp_sim.h
@@ -3831,7 +3831,7 @@ public:
     void Dump(FILE *fd);
     void DumpCsv(FILE *fd);
     void DumpPktSize();
-    void Update();
+    void UpdateFast();
     double GetCpuUtil();
 
 public:
diff --git a/src/latency.cpp b/src/latency.cpp
index fff7935d..0303e89a 100644
--- a/src/latency.cpp
+++ b/src/latency.cpp
@@ -689,17 +689,17 @@ void  CLatencyManager::try_rx(){
     for (i=0; i<m_max_ports; i++) {
         CLatencyManagerPerPort * lp=&m_ports[i];
         rte_mbuf_t * m;
-        m_cpu_dp_u.start_work();
         /* try to read 64 packets clean up the queue */
         uint16_t cnt_p = lp->m_io->rx_burst(rx_pkts, 64);
         if (cnt_p) {
+            m_cpu_dp_u.start_work1();
             int j;
             for (j=0; j<cnt_p; j++) {
                 m=rx_pkts[j] ;
                 handle_rx_pkt(lp,m);
             }
             /* commit only if there was work to do ! */
-            m_cpu_dp_u.commit();
+            m_cpu_dp_u.commit1();
           }/* if work */
       }// all ports
 }
@@ -762,12 +762,12 @@ void  CLatencyManager::start(int iter) {
 
             break;
         case CGenNode::FLOW_PKT:
-            m_cpu_dp_u.start_work();
+            m_cpu_dp_u.start_work1();
             send_pkt_all_ports();
             m_p_queue.pop();
             node->m_time += m_delta_sec;
             m_p_queue.push(node);
-            m_cpu_dp_u.commit();
+            m_cpu_dp_u.commit1();
             break;
         }
 
@@ -911,8 +911,12 @@ void CLatencyManager::rx_check_dump_json(std::string & json){
     }
 }
 
-void CLatencyManager::update(){
+
+void CLatencyManager::update_fast(){
     m_cpu_cp_u.Update() ;
+}
+
+void CLatencyManager::update(){
     for (int i=0; i<m_max_ports; i++) {
         CLatencyManagerPerPort * lp=&m_ports[i];
         lp->m_port.m_hist.update();
diff --git a/src/latency.h b/src/latency.h
index f5f90cf9..259bb221 100644
--- a/src/latency.h
+++ b/src/latency.h
@@ -356,6 +356,8 @@ public:
         return ( m_pkt_gen.get_payload_offset() );
     }
     void update();
+    void update_fast();
+
     void dump_json(std::string & json ); // dump to json 
     void dump_json_v2(std::string & json );
     void DumpRxCheckVerification(FILE *fd,uint64_t total_tx_rx_check);
diff --git a/src/main_dpdk.cpp b/src/main_dpdk.cpp
index 092bd133..fd5a1de8 100644
--- a/src/main_dpdk.cpp
+++ b/src/main_dpdk.cpp
@@ -3424,7 +3424,6 @@ void CGlobalTRex::update_stats(){
         total_open_flows +=   lpt->m_stats.m_total_open_flows ;
     }
     m_last_total_cps = m_cps.add(total_open_flows);
-    m_fl.Update();
 
 }
 
@@ -3841,6 +3840,14 @@ bool
 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();
+    }
 
     if ( is_all_cores_finished() ) {
         return false;
@@ -3861,7 +3868,7 @@ int CGlobalTRex::run_in_master() {
 
     uint32_t slow_path_counter = 0;
 
-    const int FASTPATH_DELAY_MS = 20;
+    const int FASTPATH_DELAY_MS = 10;
     const int SLOWPATH_DELAY_MS = 500;
 
     while ( true ) {
diff --git a/src/stateless/rx/trex_stateless_rx_core.cpp b/src/stateless/rx/trex_stateless_rx_core.cpp
index 2132fe9f..f7658e53 100644
--- a/src/stateless/rx/trex_stateless_rx_core.cpp
+++ b/src/stateless/rx/trex_stateless_rx_core.cpp
@@ -178,18 +178,18 @@ void CRxCoreStateless::flush_rx() {
     for (i = 0; i < m_max_ports; i++) {
         CLatencyManagerPerPort * lp = &m_ports[i];
         rte_mbuf_t * m;
-        m_cpu_dp_u.start_work();
         /* try to read 64 packets clean up the queue */
         uint16_t cnt_p = lp->m_io->rx_burst(rx_pkts, 64);
         total_pkts += cnt_p;
         if (cnt_p) {
+            m_cpu_dp_u.start_work1();
             int j;
             for (j = 0; j < cnt_p; j++) {
                 m = rx_pkts[j];
                 rte_pktmbuf_free(m);
             }
             /* commit only if there was work to do ! */
-            m_cpu_dp_u.commit();
+            m_cpu_dp_u.commit1();
         }/* if work */
     }// all ports
 }
@@ -200,11 +200,11 @@ int CRxCoreStateless::try_rx() {
     for (i = 0; i < m_max_ports; i++) {
         CLatencyManagerPerPort * lp = &m_ports[i];
         rte_mbuf_t * m;
-        m_cpu_dp_u.start_work();
         /* try to read 64 packets clean up the queue */
         uint16_t cnt_p = lp->m_io->rx_burst(rx_pkts, 64);
         total_pkts += cnt_p;
         if (cnt_p) {
+            m_cpu_dp_u.start_work1();
             int j;
             for (j = 0; j < cnt_p; j++) {
                 m = rx_pkts[j];
@@ -212,7 +212,7 @@ int CRxCoreStateless::try_rx() {
                 rte_pktmbuf_free(m);
             }
             /* commit only if there was work to do ! */
-            m_cpu_dp_u.commit();
+            m_cpu_dp_u.commit1();
         }/* if work */
     }// all ports
     return total_pkts;
@@ -249,7 +249,11 @@ void CRxCoreStateless::set_working_msg_ack(bool val) {
     sanb_smp_memory_barrier();
 }
 
-double CRxCoreStateless::get_cpu_util() {
+
+void CRxCoreStateless::update_cpu_util(){
     m_cpu_cp_u.Update();
+}
+
+double CRxCoreStateless::get_cpu_util() {
     return m_cpu_cp_u.GetVal();
 }
diff --git a/src/stateless/rx/trex_stateless_rx_core.h b/src/stateless/rx/trex_stateless_rx_core.h
index b78256c2..a372a578 100644
--- a/src/stateless/rx/trex_stateless_rx_core.h
+++ b/src/stateless/rx/trex_stateless_rx_core.h
@@ -57,6 +57,8 @@ class CRxCoreStateless {
     bool is_working() const {return (m_ack_start_work_msg == true);}
     void set_working_msg_ack(bool val);
     double get_cpu_util();
+    void update_cpu_util();
+
 
  private:
     void handle_cp_msg(TrexStatelessCpToRxMsgBase *msg);
diff --git a/src/utl_cpuu.cpp b/src/utl_cpuu.cpp
index 0701b845..4d75cf6f 100755
--- a/src/utl_cpuu.cpp
+++ b/src/utl_cpuu.cpp
@@ -26,8 +26,8 @@ limitations under the License.
 void CCpuUtlCp::Create(CCpuUtlDp * cdp){
     m_dpcpu=cdp;
     m_cpu_util=0.0;
-    m_last_total_cycles=m_dpcpu->get_total_cycles();
-    m_last_work_cycles  =m_dpcpu->get_work_cycles();
+    m_ticks=0;
+    m_work=0;
 }
 
 void CCpuUtlCp::Delete(){
@@ -36,18 +36,18 @@ void CCpuUtlCp::Delete(){
 
 
 void CCpuUtlCp::Update(){
-    uint64_t t=m_dpcpu->get_total_cycles();
-    uint64_t w=m_dpcpu->get_work_cycles();
-    uint32_t acc_total_cycles = (uint32_t)(t - m_last_total_cycles);
-	uint32_t acc_work_cycles = (uint32_t)(w - m_last_work_cycles);
-
-    m_last_total_cycles = t ; 
-    m_last_work_cycles  = w;
-
-    double window_cpu_u = ((double)acc_work_cycles/(double)acc_total_cycles);
-    
-    /* LPF*/
-    m_cpu_util = (m_cpu_util*0.75)+(window_cpu_u*0.25);
+    m_ticks++ ;
+    if ( m_dpcpu->sample_data() ) {
+        m_work++;
+    }
+    if (m_ticks==100) {
+        double window_cpu_u = ((double)m_work/(double)m_ticks);
+        /* LPF*/
+        m_cpu_util = (m_cpu_util*0.75)+(window_cpu_u*0.25);
+        m_ticks=0;
+        m_work=0;
+
+    }
 }
 
 /* return cpu % */
diff --git a/src/utl_cpuu.h b/src/utl_cpuu.h
index bb3fe67c..e5305783 100755
--- a/src/utl_cpuu.h
+++ b/src/utl_cpuu.h
@@ -29,7 +29,6 @@ class CCpuUtlDp {
 
 public:
     CCpuUtlDp(){
-        m_total_cycles=0;
         m_data=0;
     }
     inline void  start_work1(){
@@ -40,26 +39,13 @@ public:
         m_data=0;
     }
 
-    inline void start_work(){
-        m_data=os_get_hr_tick_64();
-    }
-    inline void revert(){
-    }
-    inline void commit(){
-        m_total_cycles+=(os_get_hr_tick_64()-m_data);
-    }
-    inline uint64_t get_total_cycles(void){
-        return ( os_get_hr_tick_64());
+    inline uint8_t sample_data(){
+        return (m_data);
     }
 
-    inline uint64_t get_work_cycles(void){
-        return ( m_total_cycles );
-    }
 
 private:
-    uint64_t m_total_cycles;
     uint8_t m_data;
-
 } __rte_cache_aligned; 
 
 class CCpuUtlCp {
@@ -73,12 +59,10 @@ public:
 
 private:
     CCpuUtlDp * m_dpcpu;
-    double      m_cpu_util;
-    uint64_t    m_last_total_cycles;
-    uint64_t    m_last_work_cycles;
+    uint16_t    m_ticks;
+    uint16_t    m_work;
 
-
-    // add filter 
+    double      m_cpu_util;
 };
 
 #endif