diff --git a/src/vnet/tcp/tcp.c b/src/vnet/tcp/tcp.c
index c30a693..938a863 100644
--- a/src/vnet/tcp/tcp.c
+++ b/src/vnet/tcp/tcp.c
@@ -681,7 +681,7 @@
   tc->snd_una = tc->iss;
   tc->snd_nxt = tc->iss + 1;
   tc->snd_una_max = tc->snd_nxt;
-  tc->srtt = 100;		/* 100 ms */
+  tc->srtt = 0.1 * THZ;		/* 100 ms */
 
   if (!tcp_cfg.csum_offload)
     tc->cfg_flags |= TCP_CFG_F_NO_CSUM_OFFLOAD;
@@ -1361,11 +1361,6 @@
     pool_init_fixed (tm->half_open_connections,
 		     tcp_cfg.preallocated_half_open_connections);
 
-  /* Initialize clocks per tick for TCP timestamp. Used to compute
-   * monotonically increasing timestamps. */
-  tm->tstamp_ticks_per_clock = vm->clib_time.seconds_per_clock
-    / TCP_TSTAMP_RESOLUTION;
-
   if (num_threads > 1)
     {
       clib_spinlock_init (&tm->half_open_lock);
diff --git a/src/vnet/tcp/tcp.h b/src/vnet/tcp/tcp.h
index 91783a6..bc6e353 100644
--- a/src/vnet/tcp/tcp.h
+++ b/src/vnet/tcp/tcp.h
@@ -91,7 +91,7 @@
   /** convenience pointer to this thread's vlib main */
   vlib_main_t *vm;
 
-  /** worker time */
+  /** Time measured in @ref TCP_TSTAMP_TICK used for time stamps */
   u32 time_now;
 
   /* Max timers to be handled per dispatch loop */
@@ -209,8 +209,6 @@
   /* Pool of listeners. */
   tcp_connection_t *listener_pool;
 
-  f64 tstamp_ticks_per_clock;
-
   /** vlib buffer size */
   u32 bytes_per_buffer;
 
diff --git a/src/vnet/tcp/tcp_cli.c b/src/vnet/tcp/tcp_cli.c
index b76b404..6030440 100644
--- a/src/vnet/tcp/tcp_cli.c
+++ b/src/vnet/tcp/tcp_cli.c
@@ -205,10 +205,11 @@
 	      tc->rcv_opts.tsecr, tc->tsecr_last_ack,
 	      tcp_time_now () - tc->tsval_recent_age);
   s = format (s, " snd_mss %u\n", tc->snd_mss);
-  s = format (s, " rto %u rto_boff %u srtt %u us %.3f rttvar %u rtt_ts %.4f",
-	      tc->rto, tc->rto_boff, tc->srtt, tc->mrtt_us * 1000, tc->rttvar,
-	      tc->rtt_ts);
-  s = format (s, " rtt_seq %u\n", tc->rtt_seq - tc->iss);
+  s = format (s, " rto %u rto_boff %u srtt %.1f us %.3f rttvar %.1f",
+	      tc->rto / 1000, tc->rto_boff, tc->srtt / 1000.0,
+	      tc->mrtt_us * 1e3, tc->rttvar / 1000.0);
+  s = format (s, " rtt_ts %.4f rtt_seq %u\n", tc->rtt_ts,
+	      tc->rtt_seq - tc->iss);
   s = format (s, " next_node %u opaque 0x%x fib_index %u\n",
 	      tc->next_node_index, tc->next_node_opaque, tc->c_fib_index);
   s = format (s, " cong:   %U", format_tcp_congestion, tc);
diff --git a/src/vnet/tcp/tcp_inlines.h b/src/vnet/tcp/tcp_inlines.h
index c4b155a..cb00ca4 100644
--- a/src/vnet/tcp/tcp_inlines.h
+++ b/src/vnet/tcp/tcp_inlines.h
@@ -218,9 +218,7 @@
 always_inline u32
 tcp_set_time_now (tcp_worker_ctx_t * wrk)
 {
-  tcp_main_t *tm = &tcp_main;
-  wrk->time_now = (u64) (clib_cpu_time_now () * tm->tstamp_ticks_per_clock);
-  return wrk->time_now;
+  return wrk->time_now = (u64) (vlib_time_now (wrk->vm) * TCP_TSTP_HZ);
 }
 
 always_inline tcp_connection_t *
diff --git a/src/vnet/tcp/tcp_input.c b/src/vnet/tcp/tcp_input.c
index 519219e..5fa7bf2 100644
--- a/src/vnet/tcp/tcp_input.c
+++ b/src/vnet/tcp/tcp_input.c
@@ -424,31 +424,26 @@
 /**
  * Compute smoothed RTT as per VJ's '88 SIGCOMM and RFC6298
  *
- * Note that although the original article, srtt and rttvar are scaled
+ * Note that although in the original article srtt and rttvar are scaled
  * to minimize round-off errors, here we don't. Instead, we rely on
  * better precision time measurements.
+ *
+ * A known limitation of the algorithm is that a drop in rtt results in a
+ * rttvar increase and bigger RTO.
+ *
+ * mrtt must be provided in @ref TCP_TICK multiples, i.e., in us. Note that
+ * timestamps are measured as ms ticks so they must be converted before
+ * calling this function.
  */
 static void
 tcp_estimate_rtt (tcp_connection_t * tc, u32 mrtt)
 {
   int err, diff;
 
-  if (tc->srtt != 0)
-    {
-      err = mrtt - tc->srtt;
-
-      /* XXX Drop in RTT results in RTTVAR increase and bigger RTO.
-       * The increase should be bound */
-      tc->srtt = clib_max ((int) tc->srtt + (err >> 3), 1);
-      diff = (clib_abs (err) - (int) tc->rttvar) >> 2;
-      tc->rttvar = clib_max ((int) tc->rttvar + diff, 1);
-    }
-  else
-    {
-      /* First measurement. */
-      tc->srtt = mrtt;
-      tc->rttvar = mrtt >> 1;
-    }
+  err = mrtt - tc->srtt;
+  tc->srtt = clib_max ((int) tc->srtt + (err >> 3), 1);
+  diff = (clib_abs (err) - (int) tc->rttvar) >> 2;
+  tc->rttvar = clib_max ((int) tc->rttvar + diff, 1);
 }
 
 static inline void
@@ -506,8 +501,8 @@
    * seq_lt (tc->snd_una, ack). This is a condition for calling update_rtt */
   else if (tcp_opts_tstamp (&tc->rcv_opts) && tc->rcv_opts.tsecr)
     {
-      u32 now = tcp_tstamp (tc);
-      mrtt = clib_max (now - tc->rcv_opts.tsecr, 1);
+      mrtt = clib_max (tcp_tstamp (tc) - tc->rcv_opts.tsecr, 1);
+      mrtt *= TCP_TSTP_TO_HZ;
     }
 
 estimate_rtt:
@@ -543,8 +538,8 @@
     }
   else
     {
-      mrtt = tcp_time_now_w_thread (thread_index) - tc->rcv_opts.tsecr;
-      mrtt = clib_max (mrtt, 1);
+      mrtt = tcp_tstamp (tc) - tc->rcv_opts.tsecr;
+      mrtt = clib_max (mrtt, 1) * TCP_TSTP_TO_HZ;
       /* Due to retransmits we don't know the initial mrtt */
       if (tc->rto_boff && mrtt > 1 * THZ)
 	mrtt = 1 * THZ;
@@ -552,7 +547,11 @@
     }
 
   if (mrtt > 0 && mrtt < TCP_RTT_MAX)
-    tcp_estimate_rtt (tc, mrtt);
+    {
+      /* First measurement as per RFC 6298 */
+      tc->srtt = mrtt;
+      tc->rttvar = mrtt >> 1;
+    }
   tcp_update_rto (tc);
 }
 
diff --git a/src/vnet/tcp/tcp_types.h b/src/vnet/tcp/tcp_types.h
index 3cf4e9e..d7bcac5 100644
--- a/src/vnet/tcp/tcp_types.h
+++ b/src/vnet/tcp/tcp_types.h
@@ -22,10 +22,14 @@
 #include <vnet/session/transport.h>
 #include <vppinfra/tw_timer_16t_2w_512sl.h>
 
-#define TCP_TICK 0.001			/**< TCP tick period (s) */
-#define THZ (u32) (1/TCP_TICK)		/**< TCP tick frequency */
-#define TCP_TSTAMP_RESOLUTION TCP_TICK	/**< Time stamp resolution */
-#define TCP_PAWS_IDLE 24 * 24 * 60 * 60 * THZ /**< 24 days */
+#define TCP_TICK 0.000001			/**< TCP tick period (s) */
+#define THZ (u32) (1/TCP_TICK)			/**< TCP tick frequency */
+
+#define TCP_TSTP_TICK 0.001			/**< Timestamp tick (s) */
+#define TCP_TSTP_HZ (u32) (1/TCP_TSTP_TICK)	/**< Timestamp freq */
+#define TCP_PAWS_IDLE (24 * 86400 * TCP_TSTP_HZ)/**< 24 days */
+#define TCP_TSTP_TO_HZ (u32) (TCP_TSTP_TICK * THZ)
+
 #define TCP_FIB_RECHECK_PERIOD	1 * THZ	/**< Recheck every 1s */
 #define TCP_MAX_OPTION_SPACE 40
 #define TCP_CC_DATA_SZ 24
@@ -355,7 +359,7 @@
   /* RTT and RTO */
   u32 rto;		/**< Retransmission timeout */
   u32 rto_boff;		/**< Index for RTO backoff */
-  u32 srtt;		/**< Smoothed RTT */
+  u32 srtt;		/**< Smoothed RTT measured in @ref TCP_TICK */
   u32 rttvar;		/**< Smoothed mean RTT difference. Approximates variance */
   u32 rtt_seq;		/**< Sequence number for tracked ACK */
   f64 rtt_ts;		/**< Timestamp for tracked ACK */