diff --git a/a8/perfmonitor.h b/a8/perfmonitor.h
index 9d2db44..c9576b6 100644
--- a/a8/perfmonitor.h
+++ b/a8/perfmonitor.h
@@ -1,5 +1,7 @@
 #pragma once
 
+#include <atomic>
+
 #include <a8/singleton.h>
 
 namespace a8
@@ -20,6 +22,10 @@ namespace a8
         long long recv_eagain_times = 0;
         long long max_send_time = 0;
         long long max_recv_time = 0;
+        std::atomic<long long> server_send_bytes = {0};
+        std::atomic<long long> server_consume_bytes = {0};
+        std::atomic<long long> conn_send_bytes = {0};
+        std::atomic<long long> conn_consume_bytes = {0};
 
     };
 
diff --git a/a8/tcpclient.cc b/a8/tcpclient.cc
index f91e604..665803c 100644
--- a/a8/tcpclient.cc
+++ b/a8/tcpclient.cc
@@ -13,6 +13,7 @@
 
 #include <a8/a8.h>
 #include <a8/tcpclient.h>
+#include <a8/perfmonitor.h>
 
 const int MAX_RECV_BUFFERSIZE = 1024 * 64;
 
@@ -65,6 +66,9 @@ namespace a8
     void TcpClient::SendBuff(const char* buff, unsigned int bufflen)
     {
         if (bufflen > 0) {
+#ifdef A8_PERFT
+            PerfMonitor::Instance()->conn_send_bytes += bufflen;
+#endif
             a8::SendQueueNode* p = (a8::SendQueueNode*)malloc(sizeof(a8::SendQueueNode));
             memset(p, 0, sizeof(SendQueueNode));
             p->buff = (char*)malloc(bufflen);
@@ -217,6 +221,9 @@ namespace a8
                                      currnode->buff + currnode->sent_bytes,
                                      currnode->bufflen - currnode->sent_bytes,
                                      0);
+#ifdef A8_PERFT
+                    PerfMonitor::Instance()->conn_send_bytes += len;
+#endif
                     if (len > 0) {
                         currnode->sent_bytes += len;
                     } else {
diff --git a/a8/tcpsession.cc b/a8/tcpsession.cc
index 24fb1fb..009a84b 100644
--- a/a8/tcpsession.cc
+++ b/a8/tcpsession.cc
@@ -68,6 +68,9 @@ namespace a8
             return;
         }
         if (bufflen > 0) {
+#ifdef A8_PERFT
+            PerfMonitor::Instance()->conn_send_bytes += bufflen;
+#endif
             a8::SendQueueNode* p = (a8::SendQueueNode*)malloc(sizeof(a8::SendQueueNode));
             memset(p, 0, sizeof(a8::SendQueueNode));
             p->buff = (char*)malloc(bufflen);
@@ -299,6 +302,9 @@ namespace a8
                 }
             }
             work_node_->sent_bytes += sentbytes;
+#ifdef A8_PERFT
+            PerfMonitor::Instance()->server_consume_bytes += sentbytes;
+#endif
             if (work_node_->sent_bytes >= work_node_->bufflen) {
                 a8::SendQueueNode *pdelnode = work_node_;
                 work_node_ = work_node_->next;