diff --git a/sbin/ipfw/ipfw.8 b/sbin/ipfw/ipfw.8
index 078d01014a61..52d2d3d1437c 100644
--- a/sbin/ipfw/ipfw.8
+++ b/sbin/ipfw/ipfw.8
@@ -1,7 +1,7 @@
 .\"
 .\" $FreeBSD$
 .\"
-.Dd October 19, 2007
+.Dd November 17, 2007
 .Dt IPFW 8
 .Os
 .Sh NAME
@@ -1756,6 +1756,16 @@ Depending on local policies, a flow can contain packets for a single
 TCP connection, or from/to a given host, or entire subnet, or a
 protocol type, etc.
 .Pp
+There are two modes of dummynet operation: normal and fast.
+Normal mode tries to emulate real link: dummynet scheduler ensures packet will
+not leave pipe faster than it would be on real link with given bandwidth.
+Fast mode allows certain packets to bypass dummynet scheduler (if packet flow
+does not exceed pipe's bandwidth). Thus fast mode requires less cpu cycles
+per packet (in average) but packet latency can be significantly lower comparing
+to real link with same bandwidth. Default is normal mode, fast mode can be
+enabled by setting net.inet.ip.dummynet.io_fast sysctl(8) variable to non-zero
+value.
+.Pp
 Packets belonging to the same flow are then passed to either of two
 different objects, which implement the traffic regulation:
 .Bl -hang -offset XXXX
@@ -2120,6 +2130,14 @@ Default size of the hash table used for dynamic pipes/queues.
 This value is used when no
 .Cm buckets
 option is specified when configuring a pipe/queue.
+.It Em net.inet.ip.dummynet.io_fast : No 0
+If set to non-zero value enables "fast" mode of dummynet operation (see above).
+.It Em net.inet.ip.dummynet.io_pkt
+Number of packets passed to by dummynet.
+.It Em net.inet.ip.dummynet.io_pkt_drop
+Number of packets dropped by dummynet.
+.It Em net.inet.ip.dummynet.io_pkt_fast
+Number of packets bypassed dummynet scheduler.
 .It Em net.inet.ip.dummynet.max_chain_len : No 16
 Target value for the maximum number of pipes/queues in a hash bucket.
 The product
diff --git a/sys/netinet/ip_dummynet.c b/sys/netinet/ip_dummynet.c
index d6f1880a5375..cb079a5443b1 100644
--- a/sys/netinet/ip_dummynet.c
+++ b/sys/netinet/ip_dummynet.c
@@ -110,6 +110,7 @@ static long tick_lost;			/* Lost(coalesced) ticks number. */
 /* Adjusted vs non-adjusted curr_time difference (ticks). */
 static long tick_diff;
 
+static int		io_fast;
 static unsigned long	io_pkt;
 static unsigned long	io_pkt_fast;
 static unsigned long	io_pkt_drop;
@@ -185,6 +186,8 @@ SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_diff,
 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_lost,
     CTLFLAG_RD, &tick_lost, 0,
     "Number of ticks coalesced by dummynet taskqueue.");
+SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, io_fast,
+    CTLFLAG_RW, &io_fast, 0, "Enable fast dummynet io.");
 SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt,
     CTLFLAG_RD, &io_pkt, 0,
     "Number of packets passed to dummynet.");
@@ -967,7 +970,7 @@ create_queue(struct dn_flow_set *fs, int i)
 	q->hash_slot = i;
 	q->next = fs->rq[i];
 	q->S = q->F + 1;	/* hack - mark timestamp as invalid. */
-	q->numbytes = fs->pipe->bandwidth;
+	q->numbytes = io_fast ? fs->pipe->bandwidth : 0;
 	fs->rq[i] = q;
 	fs->rq_elements++;
 	return (q);
@@ -1325,7 +1328,7 @@ dummynet_io(struct mbuf **m0, int dir, struct ip_fw_args *fwa)
 		goto done;
 
 	if (q->q_time < curr_time)
-		q->numbytes = fs->pipe->bandwidth;
+		q->numbytes = io_fast ? fs->pipe->bandwidth : 0;
 	q->q_time = curr_time;
 
 	/*
@@ -1736,7 +1739,7 @@ config_pipe(struct dn_pipe *p)
 			/* Flush accumulated credit for all queues. */
 			for (i = 0; i <= pipe->fs.rq_size; i++)
 				for (q = pipe->fs.rq[i]; q; q = q->next)
-					q->numbytes = p->bandwidth;
+					q->numbytes = io_fast ? p->bandwidth : 0;
 
 		pipe->bandwidth = p->bandwidth;
 		pipe->numbytes = 0;		/* just in case... */