Skip to contentSkip to conversation

net: tcp: RTO restart

PATCH In review Per Hurtig 7 participants 17 messages Feb 18, 2014 – Feb 26, 2014

Subsystems (3)
lore
Share & tools
lore

Per Hurtig

author
#0 lore raw

PATCH 1/1 net: tcp: RTO restart

This patch implements the RTO restart modification described in
http://tools.ietf.org/html/draft-ietf-tcpm-rtorestart-02

RTO Restart's goal is to provide quicker loss recovery for segments lost in the end of a burst/connection. To accomplish this the algorithm adjusts the RTO value on each rearm of the retransmission timer to be exactly RTO ms after the earliest outstanding segment was sent. The offsetting against the earliest outstanding segment is not done by the regular rearm algorithm, which causes RTOs to occur, on average, after RTO+RTT ms.

As a faster timeout is only beneficial in scenarios where fast retransmit/early retransmit cannot be triggered the algorithm will only kick in when there is a small amount of outstanding segments.

The RTO Restart proposal is accepted as a working group item in the IETF TCP Maintenance and Minor Extensions (tcpm) TCP wg and is intended for experimental RFC status.

---
 include/net/tcp.h          |    1 +
 net/ipv4/sysctl_net_ipv4.c |    7 +++++++
 net/ipv4/tcp_input.c       |   11 +++++++++++
 3 files changed, 19 insertions(+)
include/net/tcp.h +1 -0
index 56fc366..575e82a 100644
@@ -278,6 +278,7 @@ extern int sysctl_tcp_slow_start_after_idle;
 extern int sysctl_tcp_thin_linear_timeouts;
 extern int sysctl_tcp_thin_dupack;
 extern int sysctl_tcp_early_retrans;
+extern int sysctl_tcp_rto_restart;
 extern int sysctl_tcp_limit_output_bytes;
 extern int sysctl_tcp_challenge_ack_limit;
 extern unsigned int sysctl_tcp_notsent_lowat;
net/ipv4/sysctl_net_ipv4.c +7 -0
index 44eba05..c605f4f 100644
@@ -717,6 +717,13 @@ static struct ctl_table ipv4_table[] = {
 		.extra2		= &four,
 	},
 	{
+		.procname	= "tcp_rto_restart",
+		.data		= &sysctl_tcp_rto_restart,
+		.maxlen		= sizeof(int),
+		.mode		= 0644,
+		.proc_handler	= proc_dointvec,
+	},
+	{
 		.procname	= "tcp_min_tso_segs",
 		.data		= &sysctl_tcp_min_tso_segs,
 		.maxlen		= sizeof(int),
net/ipv4/tcp_input.c +11 -0
index 227cba7..450ee30 100644
@@ -98,6 +98,7 @@ int sysctl_tcp_thin_dupack __read_mostly;
 
 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
 int sysctl_tcp_early_retrans __read_mostly = 3;
+int sysctl_tcp_rto_restart __read_mostly = 1;
 
 #define FLAG_DATA		0x01 /* Incoming frame contained data.		*/
 #define FLAG_WIN_UPDATE		0x02 /* Incoming ACK was a window update.	*/
@@ -2972,6 +2973,16 @@ void tcp_rearm_rto(struct sock *sk)
 			 */
 			if (delta > 0)
 				rto = delta;
+		} else if (icsk->icsk_pending == ICSK_TIME_RETRANS &&
+			   sysctl_tcp_rto_restart &&
+			   sk->sk_send_head == NULL &&
+			   tp->packets_out < 4) {
+			struct sk_buff *skb = tcp_write_queue_head(sk);
+			const u32 rto_time_stamp = TCP_SKB_CB(skb)->when;
+			s32 delta = (s32)(tcp_time_stamp - rto_time_stamp);
+
+			if (delta > 0)
+				rto -= delta;
 		}
 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
 					  TCP_RTO_MAX);
Signature · 1 line
1.7.9.5
#1 lore raw
Quoted 10 lines · Per Hurtig

This patch implements the RTO restart modification described in
http://tools.ietf.org/html/draft-ietf-tcpm-rtorestart-02

RTO Restart's goal is to provide quicker loss recovery for segments lost in the end of a burst/connection. To accomplish this the algorithm adjusts the RTO value on each rearm of the retransmission timer to be exactly RTO ms after the earliest outstanding segment was sent. The offsetting against the earliest outstanding segment is not done by the regular rearm algorithm, which causes RTOs to occur, on average, after RTO+RTT ms.

Quoted 3 lines

As a faster timeout is only beneficial in scenarios where fast retransmit/early retransmit cannot be triggered the algorithm will only kick in when there is a small amount of outstanding segments.

(repost in plaine-text, sorry for the duplication)

I am not sure this works well with Linux min-RTO=200ms, and don't feel comfortable this is default on without some real experiments.

I've implemented (a nearly identical version of) rto-restart on Google web servers and found #timeouts increased by ~30%. Although the fast timeout helps really short flows, it has a very negative side-effect:
resetting cwnd to 1. Thus the next HTTP response may start with a small cwnd. In contrast, TCP loss probe has fast timeout (1.5RTT) to do fast recovery to avoid the side-effect. In other words, I am doubtful we need rto-restart with TCP loss probe, but applying RTO-restart on TLP timer may be useful.

I've voiced this concern multiple times in ietf tcpm discussion when we discuss this draft: that the idea itself is fine, but we'll need to change Linux RTO algorithm together, not just the timer itself.

I am happy to post some more detailed data if people are interested.

Quoted 78 lines

The RTO Restart proposal is accepted as a working group item in the IETF TCP Maintenance and Minor Extensions (tcpm) TCP wg and is intended for experimental RFC status.

---
 include/net/tcp.h          |    1 +
 net/ipv4/sysctl_net_ipv4.c |    7 +++++++
 net/ipv4/tcp_input.c       |   11 +++++++++++
 3 files changed, 19 insertions(+)
include/net/tcp.h +1 -0
index 56fc366..575e82a 100644
@@ -278,6 +278,7 @@ extern int sysctl_tcp_slow_start_after_idle;
 extern int sysctl_tcp_thin_linear_timeouts;
 extern int sysctl_tcp_thin_dupack;
 extern int sysctl_tcp_early_retrans;
+extern int sysctl_tcp_rto_restart;
 extern int sysctl_tcp_limit_output_bytes;
 extern int sysctl_tcp_challenge_ack_limit;
 extern unsigned int sysctl_tcp_notsent_lowat;
net/ipv4/sysctl_net_ipv4.c +7 -0
index 44eba05..c605f4f 100644
@@ -717,6 +717,13 @@ static struct ctl_table ipv4_table[] = {
                .extra2         = &four,
        },
        {
+               .procname       = "tcp_rto_restart",
+               .data           = &sysctl_tcp_rto_restart,
+               .maxlen         = sizeof(int),
+               .mode           = 0644,
+               .proc_handler   = proc_dointvec,
+       },
+       {
                .procname       = "tcp_min_tso_segs",
                .data           = &sysctl_tcp_min_tso_segs,
                .maxlen         = sizeof(int),
net/ipv4/tcp_input.c +11 -1
index 227cba7..450ee30 100644
@@ -98,6 +98,7 @@ int sysctl_tcp_thin_dupack __read_mostly;
 
 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
 int sysctl_tcp_early_retrans __read_mostly = 3;
+int sysctl_tcp_rto_restart __read_mostly = 1;
 
 #define FLAG_DATA              0x01 /* Incoming frame contained data.          */
 #define FLAG_WIN_UPDATE                0x02 /* Incoming ACK was a window update.       */
@@ -2972,6 +2973,16 @@ void tcp_rearm_rto(struct sock *sk)
                         */
                        if (delta > 0)
                                rto = delta;
+               } else if (icsk->icsk_pending == ICSK_TIME_RETRANS &&
+                          sysctl_tcp_rto_restart &&
+                          sk->sk_send_head == NULL &&
+                          tp->packets_out < 4) {
+                       struct sk_buff *skb = tcp_write_queue_head(sk);
+                       const u32 rto_time_stamp = TCP_SKB_CB(skb)->when;
+                       s32 delta = (s32)(tcp_time_stamp - rto_time_stamp);
+
+                       if (delta > 0)
+                               rto -= delta;
                }
                inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
                                          TCP_RTO_MAX);
--

1.7.9.5

--
To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

author
#2 lore raw

Hi Yuchung,

see inline

Quoted 19 lines · Yuchung Cheng
Quoted 10 lines · Per Hurtig

This patch implements the RTO restart modification described in
http://tools.ietf.org/html/draft-ietf-tcpm-rtorestart-02

RTO Restart's goal is to provide quicker loss recovery for segments lost in the end of a burst/connection. To accomplish this the algorithm adjusts the RTO value on each rearm of the retransmission timer to be exactly RTO ms after the earliest outstanding segment was sent. The offsetting against the earliest outstanding segment is not done by the regular rearm algorithm, which causes RTOs to occur, on average, after RTO+RTT ms.

Quoted 3 lines

As a faster timeout is only beneficial in scenarios where fast retransmit/early retransmit cannot be triggered the algorithm will only kick in when there is a small amount of outstanding segments.

(repost in plaine-text, sorry for the duplication)

I am not sure this works well with Linux min-RTO=200ms, and don't feel comfortable this is default on without some real experiments.

Quoted 8 lines

I've implemented (a nearly identical version of) rto-restart on Google web servers and found #timeouts increased by ~30%. Although the fast timeout helps really short flows, it has a very negative side-effect:
resetting cwnd to 1. Thus the next HTTP response may start with a small cwnd. In contrast, TCP loss probe has fast timeout (1.5RTT) to do fast recovery to avoid the side-effect. In other words, I am doubtful we need rto-restart with TCP loss probe, but applying RTO-restart on TLP timer may be useful.

As we have discussed, I agree that RTO restart can also be applied to TLP and this is also mentioned in the ietf draft. However, I think they are also complementary in reducing loss recovery delay.

Quoted 6 lines

I've voiced this concern multiple times in ietf tcpm discussion when we discuss this draft: that the idea itself is fine, but we'll need to change Linux RTO algorithm together, not just the timer itself.

I am happy to post some more detailed data if people are interested.

We would love to see the data. The last time we discussed this, you were not able to find it. From the latest discussions in the tcpm group I understood you didn't think it was that big of a problem anymore.

Best Regards,
Per Hurtig

Quoted 83 lines
Quoted 75 lines

The RTO Restart proposal is accepted as a working group item in the IETF TCP Maintenance and Minor Extensions (tcpm) TCP wg and is intended for experimental RFC status.

---
 include/net/tcp.h          |    1 +
 net/ipv4/sysctl_net_ipv4.c |    7 +++++++
 net/ipv4/tcp_input.c       |   11 +++++++++++
 3 files changed, 19 insertions(+)
include/net/tcp.h +1 -0
index 56fc366..575e82a 100644
@@ -278,6 +278,7 @@ extern int sysctl_tcp_slow_start_after_idle;
 extern int sysctl_tcp_thin_linear_timeouts;
 extern int sysctl_tcp_thin_dupack;
 extern int sysctl_tcp_early_retrans;
+extern int sysctl_tcp_rto_restart;
 extern int sysctl_tcp_limit_output_bytes;
 extern int sysctl_tcp_challenge_ack_limit;
 extern unsigned int sysctl_tcp_notsent_lowat;
net/ipv4/sysctl_net_ipv4.c +7 -0
index 44eba05..c605f4f 100644
@@ -717,6 +717,13 @@ static struct ctl_table ipv4_table[] = {
                .extra2         = &four,
        },
        {
+               .procname       = "tcp_rto_restart",
+               .data           = &sysctl_tcp_rto_restart,
+               .maxlen         = sizeof(int),
+               .mode           = 0644,
+               .proc_handler   = proc_dointvec,
+       },
+       {
                .procname       = "tcp_min_tso_segs",
                .data           = &sysctl_tcp_min_tso_segs,
                .maxlen         = sizeof(int),
net/ipv4/tcp_input.c +11 -1
index 227cba7..450ee30 100644
@@ -98,6 +98,7 @@ int sysctl_tcp_thin_dupack __read_mostly;
 
 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
 int sysctl_tcp_early_retrans __read_mostly = 3;
+int sysctl_tcp_rto_restart __read_mostly = 1;
 
 #define FLAG_DATA              0x01 /* Incoming frame contained data.          */
 #define FLAG_WIN_UPDATE                0x02 /* Incoming ACK was a window update.       */
@@ -2972,6 +2973,16 @@ void tcp_rearm_rto(struct sock *sk)
                         */
                        if (delta > 0)
                                rto = delta;
+               } else if (icsk->icsk_pending == ICSK_TIME_RETRANS &&
+                          sysctl_tcp_rto_restart &&
+                          sk->sk_send_head == NULL &&
+                          tp->packets_out < 4) {
+                       struct sk_buff *skb = tcp_write_queue_head(sk);
+                       const u32 rto_time_stamp = TCP_SKB_CB(skb)->when;
+                       s32 delta = (s32)(tcp_time_stamp - rto_time_stamp);
+
+                       if (delta > 0)
+                               rto -= delta;
                }
                inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
                                          TCP_RTO_MAX);
--

1.7.9.5

Quoted 4 lines

--
To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

--
To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#3 lore raw
Quoted 38 lines · Per Hurtig

Hi Yuchung,

see inline

Quoted 19 lines · Yuchung Cheng
Quoted 10 lines · Per Hurtig

This patch implements the RTO restart modification described in
http://tools.ietf.org/html/draft-ietf-tcpm-rtorestart-02

RTO Restart's goal is to provide quicker loss recovery for segments lost in the end of a burst/connection. To accomplish this the algorithm adjusts the RTO value on each rearm of the retransmission timer to be exactly RTO ms after the earliest outstanding segment was sent. The offsetting against the earliest outstanding segment is not done by the regular rearm algorithm, which causes RTOs to occur, on average, after RTO+RTT ms.

Quoted 3 lines

As a faster timeout is only beneficial in scenarios where fast retransmit/early retransmit cannot be triggered the algorithm will only kick in when there is a small amount of outstanding segments.

(repost in plaine-text, sorry for the duplication)

I am not sure this works well with Linux min-RTO=200ms, and don't feel comfortable this is default on without some real experiments.

Quoted 8 lines

I've implemented (a nearly identical version of) rto-restart on Google web servers and found #timeouts increased by ~30%. Although the fast timeout helps really short flows, it has a very negative side-effect:
resetting cwnd to 1. Thus the next HTTP response may start with a small cwnd. In contrast, TCP loss probe has fast timeout (1.5RTT) to do fast recovery to avoid the side-effect. In other words, I am doubtful we need rto-restart with TCP loss probe, but applying RTO-restart on TLP timer may be useful.

As we have discussed, I agree that RTO restart can also be applied to TLP and this is also mentioned in the ietf draft. However, I think they are also complementary in reducing loss recovery delay.

Quoted 1 line
Quoted 7 lines
Quoted 5 lines

I've voiced this concern multiple times in ietf tcpm discussion when we discuss this draft: that the idea itself is fine, but we'll need to change Linux RTO algorithm together, not just the timer itself.

I am happy to post some more detailed data if people are interested.

We would love to see the data. The last time we discussed this, you were not

Sure I can gather some data this week for rto-restart.

Quoted 2 lines

able to find it. From the latest discussions in the tcpm group I understood you didn't think it was that big of a problem anymore

rto-restart draft is OK with the context of hypothetical stack implementing only RFC standards. But Linux is (much) more complex:
min-RTO=200ms, TLP, FACK, etc. Changing how the default timer is armed needs to be carefully tested in real world before making rto-restart default. Shortening timer (retry faster) always help short connections but the long connections may suffer the cwnd reset due to spurious timeout, *especially* Linux is already running a very aggressive min-rto.

btw, the rto -= delta may overflow if the socket was locked when the timer fired previously?

Quoted 88 lines

Best Regards,
Per Hurtig

Quoted 83 lines
Quoted 75 lines

The RTO Restart proposal is accepted as a working group item in the IETF TCP Maintenance and Minor Extensions (tcpm) TCP wg and is intended for experimental RFC status.

---
 include/net/tcp.h          |    1 +
 net/ipv4/sysctl_net_ipv4.c |    7 +++++++
 net/ipv4/tcp_input.c       |   11 +++++++++++
 3 files changed, 19 insertions(+)
include/net/tcp.h +1 -0
index 56fc366..575e82a 100644
@@ -278,6 +278,7 @@ extern int sysctl_tcp_slow_start_after_idle;
 extern int sysctl_tcp_thin_linear_timeouts;
 extern int sysctl_tcp_thin_dupack;
 extern int sysctl_tcp_early_retrans;
+extern int sysctl_tcp_rto_restart;
 extern int sysctl_tcp_limit_output_bytes;
 extern int sysctl_tcp_challenge_ack_limit;
 extern unsigned int sysctl_tcp_notsent_lowat;
net/ipv4/sysctl_net_ipv4.c +7 -0
index 44eba05..c605f4f 100644
@@ -717,6 +717,13 @@ static struct ctl_table ipv4_table[] = {
                .extra2         = &four,
        },
        {
+               .procname       = "tcp_rto_restart",
+               .data           = &sysctl_tcp_rto_restart,
+               .maxlen         = sizeof(int),
+               .mode           = 0644,
+               .proc_handler   = proc_dointvec,
+       },
+       {
                .procname       = "tcp_min_tso_segs",
                .data           = &sysctl_tcp_min_tso_segs,
                .maxlen         = sizeof(int),
net/ipv4/tcp_input.c +11 -1
index 227cba7..450ee30 100644
@@ -98,6 +98,7 @@ int sysctl_tcp_thin_dupack __read_mostly;
 
 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
 int sysctl_tcp_early_retrans __read_mostly = 3;
+int sysctl_tcp_rto_restart __read_mostly = 1;
 
 #define FLAG_DATA              0x01 /* Incoming frame contained data.          */
 #define FLAG_WIN_UPDATE                0x02 /* Incoming ACK was a window update.       */
@@ -2972,6 +2973,16 @@ void tcp_rearm_rto(struct sock *sk)
                         */
                        if (delta > 0)
                                rto = delta;
+               } else if (icsk->icsk_pending == ICSK_TIME_RETRANS &&
+                          sysctl_tcp_rto_restart &&
+                          sk->sk_send_head == NULL &&
+                          tp->packets_out < 4) {
+                       struct sk_buff *skb = tcp_write_queue_head(sk);
+                       const u32 rto_time_stamp = TCP_SKB_CB(skb)->when;
+                       s32 delta = (s32)(tcp_time_stamp - rto_time_stamp);
+
+                       if (delta > 0)
+                               rto -= delta;
                }
                inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
                                          TCP_RTO_MAX);
--

1.7.9.5

Quoted 4 lines

--
To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

--
To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

author
#4 lore raw
Quoted 20 lines · Yuchung Cheng
Quoted 19 lines · Per Hurtig

Hi Yuchung,

see inline

Quoted 14 lines · Yuchung Cheng
Quoted 13 lines · Per Hurtig

This patch implements the RTO restart modification described in
http://tools.ietf.org/html/draft-ietf-tcpm-rtorestart-02

RTO Restart's goal is to provide quicker loss recovery for segments lost in the end of a burst/connection. To accomplish this the algorithm adjusts the RTO value on each rearm of the retransmission timer to be exactly RTO ms after the earliest outstanding segment was sent. The offsetting against the earliest outstanding segment is not done by the regular rearm algorithm, which causes RTOs to occur, on average, after RTO+RTT ms.

As a faster timeout is only beneficial in scenarios where fast retransmit/early retransmit cannot be triggered the algorithm will only kick in when there is a small amount of outstanding segments.

Quoted 7 lines
Quoted 7 lines
Quoted 7 lines

(repost in plaine-text, sorry for the duplication)

I am not sure this works well with Linux min-RTO=200ms, and don't feel comfortable this is default on without some real experiments.

I've implemented (a nearly identical version of) rto-restart on Google web servers and found #timeouts increased by ~30%.

How did your version work, did it apply the modified restart for any amount of outstanding segments? If so, I’m not surprised if the #timeouts increased dramatically.

Quoted 13 lines
Quoted 13 lines
Quoted 7 lines

Although the fast
timeout helps really short flows, it has a very negative side-effect:
resetting cwnd to 1. Thus the next HTTP response may start with a small cwnd. In contrast, TCP loss probe has fast timeout (1.5RTT) to do fast recovery to avoid the side-effect. In other words, I am doubtful we need rto-restart with TCP loss probe, but applying RTO-restart on TLP timer may be useful.

As we have discussed, I agree that RTO restart can also be applied to TLP and this is also mentioned in the ietf draft. However, I think they are also complementary in reducing loss recovery delay.

Quoted 1 line
Quoted 8 lines
Quoted 7 lines
Quoted 5 lines

I've voiced this concern multiple times in ietf tcpm discussion when we discuss this draft: that the idea itself is fine, but we'll need to change Linux RTO algorithm together, not just the timer itself.

I am happy to post some more detailed data if people are interested.

We would love to see the data. The last time we discussed this, you were not

Sure I can gather some data this week for rto-restart.

That would be great! If you will be in London for the IETF perhaps we could also schedule a time to discuss it?

Quoted 12 lines
Quoted 3 lines

able to find it. From the latest discussions in the tcpm group I understood you didn't think it was that big of a problem anymore

rto-restart draft is OK with the context of hypothetical stack implementing only RFC standards. But Linux is (much) more complex:
min-RTO=200ms, TLP, FACK, etc. Changing how the default timer is armed needs to be carefully tested in real world before making rto-restart default. Shortening timer (retry faster) always help short connections but the long connections may suffer the cwnd reset due to spurious timeout, *especially* Linux is already running a very aggressive min-rto.

Quoted 2 lines

btw, the rto -= delta may overflow if the socket was locked when the timer fired previously?

Thanks for pointing that out, I think you might be right.

Best Regards,
Per Hurtig

Quoted 89 lines
Quoted 89 lines

Best Regards,
Per Hurtig

Quoted 83 lines
Quoted 75 lines

The RTO Restart proposal is accepted as a working group item in the IETF TCP Maintenance and Minor Extensions (tcpm) TCP wg and is intended for experimental RFC status.

---

include/net/tcp.h          |    1 +
net/ipv4/sysctl_net_ipv4.c |    7 +++++++
net/ipv4/tcp_input.c       |   11 +++++++++++
3 files changed, 19 insertions(+)
include/net/tcp.h +1 -0
index 56fc366..575e82a 100644
@@ -278,6 +278,7 @@ extern int sysctl_tcp_slow_start_after_idle;
 extern int sysctl_tcp_thin_linear_timeouts;
 extern int sysctl_tcp_thin_dupack;
 extern int sysctl_tcp_early_retrans;
+extern int sysctl_tcp_rto_restart;
 extern int sysctl_tcp_limit_output_bytes;
 extern int sysctl_tcp_challenge_ack_limit;
 extern unsigned int sysctl_tcp_notsent_lowat;
net/ipv4/sysctl_net_ipv4.c +7 -0
index 44eba05..c605f4f 100644
@@ -717,6 +717,13 @@ static struct ctl_table ipv4_table[] = {
               .extra2         = &four,
       },
       {
+               .procname       = "tcp_rto_restart",
+               .data           = &sysctl_tcp_rto_restart,
+               .maxlen         = sizeof(int),
+               .mode           = 0644,
+               .proc_handler   = proc_dointvec,
+       },
+       {
               .procname       = "tcp_min_tso_segs",
               .data           = &sysctl_tcp_min_tso_segs,
               .maxlen         = sizeof(int),
net/ipv4/tcp_input.c +11 -1
index 227cba7..450ee30 100644
@@ -98,6 +98,7 @@ int sysctl_tcp_thin_dupack __read_mostly;
 
 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
 int sysctl_tcp_early_retrans __read_mostly = 3;
+int sysctl_tcp_rto_restart __read_mostly = 1;
 
 #define FLAG_DATA              0x01 /* Incoming frame contained data.          */
 #define FLAG_WIN_UPDATE                0x02 /* Incoming ACK was a window update.       */
@@ -2972,6 +2973,16 @@ void tcp_rearm_rto(struct sock *sk)
                        */
                       if (delta > 0)
                               rto = delta;
+               } else if (icsk->icsk_pending == ICSK_TIME_RETRANS &&
+                          sysctl_tcp_rto_restart &&
+                          sk->sk_send_head == NULL &&
+                          tp->packets_out < 4) {
+                       struct sk_buff *skb = tcp_write_queue_head(sk);
+                       const u32 rto_time_stamp = TCP_SKB_CB(skb)->when;
+                       s32 delta = (s32)(tcp_time_stamp - rto_time_stamp);
+
+                       if (delta > 0)
+                               rto -= delta;
               }
               inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
                                         TCP_RTO_MAX);
--

1.7.9.5

Quoted 4 lines

--
To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

--
To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#5 lore raw
Quoted 53 lines · Per Hurtig
Quoted 20 lines · Yuchung Cheng
Quoted 19 lines · Per Hurtig

Hi Yuchung,

see inline

Quoted 14 lines · Yuchung Cheng
Quoted 13 lines · Per Hurtig

This patch implements the RTO restart modification described in
http://tools.ietf.org/html/draft-ietf-tcpm-rtorestart-02

RTO Restart's goal is to provide quicker loss recovery for segments lost in the end of a burst/connection. To accomplish this the algorithm adjusts the RTO value on each rearm of the retransmission timer to be exactly RTO ms after the earliest outstanding segment was sent. The offsetting against the earliest outstanding segment is not done by the regular rearm algorithm, which causes RTOs to occur, on average, after RTO+RTT ms.

As a faster timeout is only beneficial in scenarios where fast retransmit/early retransmit cannot be triggered the algorithm will only kick in when there is a small amount of outstanding segments.

Quoted 7 lines
Quoted 7 lines
Quoted 7 lines

(repost in plaine-text, sorry for the duplication)

I am not sure this works well with Linux min-RTO=200ms, and don't feel comfortable this is default on without some real experiments.

I've implemented (a nearly identical version of) rto-restart on Google web servers and found #timeouts increased by ~30%.

How did your version work, did it apply the modified restart for any amount of outstanding segments? If so, I’m not surprised if the #timeouts increased dramatically.

Quoted 17 lines
Quoted 17 lines
Quoted 7 lines

Although the fast
timeout helps really short flows, it has a very negative side-effect:
resetting cwnd to 1. Thus the next HTTP response may start with a small cwnd. In contrast, TCP loss probe has fast timeout (1.5RTT) to do fast recovery to avoid the side-effect. In other words, I am doubtful we need rto-restart with TCP loss probe, but applying RTO-restart on TLP timer may be useful.

As we have discussed, I agree that RTO restart can also be applied to TLP and this is also mentioned in the ietf draft. However, I think they are also complementary in reducing loss recovery delay.

Quoted 5 lines

I've voiced this concern multiple times in ietf tcpm discussion when we discuss this draft: that the idea itself is fine, but we'll need to change Linux RTO algorithm together, not just the timer itself.

Quoted 7 lines
Quoted 4 lines
Quoted 3 lines
Quoted 1 line

I am happy to post some more detailed data if people are interested.

We would love to see the data. The last time we discussed this, you were not

Sure I can gather some data this week for rto-restart.

That would be great! If you will be in London for the IETF perhaps we could also schedule a time to discuss it?

Sure but I should have the data by this weekend. I tweaked your patch about the rto overflow issue.

I think it's worth the effort to evaluate RTO more thoroughly and preferably use the same re-arming logic for retransmission. In other words, always offset the time elapsed since last sent when re-arming a retransmit timer, and relax the RTO calculation formula but not depending on the inflight. For example, we already use similar logic when handling ICMP dest-unreach. in tcp_ipv{4|6}.c.

Quoted 112 lines
Quoted 13 lines
Quoted 4 lines

able to find it. From the latest discussions in the tcpm group I understood you didn't think it was that big of a problem anymore

rto-restart draft is OK with the context of hypothetical stack implementing only RFC standards. But Linux is (much) more complex:
min-RTO=200ms, TLP, FACK, etc. Changing how the default timer is armed needs to be carefully tested in real world before making rto-restart default. Shortening timer (retry faster) always help short connections but the long connections may suffer the cwnd reset due to spurious timeout, *especially* Linux is already running a very aggressive min-rto.

Quoted 2 lines

btw, the rto -= delta may overflow if the socket was locked when the timer fired previously?

Thanks for pointing that out, I think you might be right.

Best Regards,
Per Hurtig

Quoted 89 lines
Quoted 89 lines

Best Regards,
Per Hurtig

Quoted 83 lines
Quoted 75 lines

The RTO Restart proposal is accepted as a working group item in the IETF TCP Maintenance and Minor Extensions (tcpm) TCP wg and is intended for experimental RFC status.

---

include/net/tcp.h          |    1 +
net/ipv4/sysctl_net_ipv4.c |    7 +++++++
net/ipv4/tcp_input.c       |   11 +++++++++++
3 files changed, 19 insertions(+)
include/net/tcp.h +1 -0
index 56fc366..575e82a 100644
@@ -278,6 +278,7 @@ extern int sysctl_tcp_slow_start_after_idle;
 extern int sysctl_tcp_thin_linear_timeouts;
 extern int sysctl_tcp_thin_dupack;
 extern int sysctl_tcp_early_retrans;
+extern int sysctl_tcp_rto_restart;
 extern int sysctl_tcp_limit_output_bytes;
 extern int sysctl_tcp_challenge_ack_limit;
 extern unsigned int sysctl_tcp_notsent_lowat;
net/ipv4/sysctl_net_ipv4.c +7 -0
index 44eba05..c605f4f 100644
@@ -717,6 +717,13 @@ static struct ctl_table ipv4_table[] = {
               .extra2         = &four,
       },
       {
+               .procname       = "tcp_rto_restart",
+               .data           = &sysctl_tcp_rto_restart,
+               .maxlen         = sizeof(int),
+               .mode           = 0644,
+               .proc_handler   = proc_dointvec,
+       },
+       {
               .procname       = "tcp_min_tso_segs",
               .data           = &sysctl_tcp_min_tso_segs,
               .maxlen         = sizeof(int),
net/ipv4/tcp_input.c +11 -1
index 227cba7..450ee30 100644
@@ -98,6 +98,7 @@ int sysctl_tcp_thin_dupack __read_mostly;
 
 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
 int sysctl_tcp_early_retrans __read_mostly = 3;
+int sysctl_tcp_rto_restart __read_mostly = 1;
 
 #define FLAG_DATA              0x01 /* Incoming frame contained data.          */
 #define FLAG_WIN_UPDATE                0x02 /* Incoming ACK was a window update.       */
@@ -2972,6 +2973,16 @@ void tcp_rearm_rto(struct sock *sk)
                        */
                       if (delta > 0)
                               rto = delta;
+               } else if (icsk->icsk_pending == ICSK_TIME_RETRANS &&
+                          sysctl_tcp_rto_restart &&
+                          sk->sk_send_head == NULL &&
+                          tp->packets_out < 4) {
+                       struct sk_buff *skb = tcp_write_queue_head(sk);
+                       const u32 rto_time_stamp = TCP_SKB_CB(skb)->when;
+                       s32 delta = (s32)(tcp_time_stamp - rto_time_stamp);
+
+                       if (delta > 0)
+                               rto -= delta;
               }
               inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
                                         TCP_RTO_MAX);
--

1.7.9.5

Quoted 4 lines

--
To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

--
To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#6 lore raw
Quoted 64 lines · Yuchung Cheng
Quoted 53 lines · Per Hurtig
Quoted 20 lines · Yuchung Cheng
Quoted 19 lines · Per Hurtig

Hi Yuchung,

see inline

Quoted 14 lines · Yuchung Cheng
Quoted 13 lines · Per Hurtig

This patch implements the RTO restart modification described in
http://tools.ietf.org/html/draft-ietf-tcpm-rtorestart-02

RTO Restart's goal is to provide quicker loss recovery for segments lost in the end of a burst/connection. To accomplish this the algorithm adjusts the RTO value on each rearm of the retransmission timer to be exactly RTO ms after the earliest outstanding segment was sent. The offsetting against the earliest outstanding segment is not done by the regular rearm algorithm, which causes RTOs to occur, on average, after RTO+RTT ms.

As a faster timeout is only beneficial in scenarios where fast retransmit/early retransmit cannot be triggered the algorithm will only kick in when there is a small amount of outstanding segments.

Quoted 7 lines
Quoted 7 lines
Quoted 7 lines

(repost in plaine-text, sorry for the duplication)

I am not sure this works well with Linux min-RTO=200ms, and don't feel comfortable this is default on without some real experiments.

I've implemented (a nearly identical version of) rto-restart on Google web servers and found #timeouts increased by ~30%.

How did your version work, did it apply the modified restart for any amount of outstanding segments? If so, I’m not surprised if the #timeouts increased dramatically.

Quoted 17 lines
Quoted 17 lines
Quoted 7 lines

Although the fast
timeout helps really short flows, it has a very negative side-effect:
resetting cwnd to 1. Thus the next HTTP response may start with a small cwnd. In contrast, TCP loss probe has fast timeout (1.5RTT) to do fast recovery to avoid the side-effect. In other words, I am doubtful we need rto-restart with TCP loss probe, but applying RTO-restart on TLP timer may be useful.

As we have discussed, I agree that RTO restart can also be applied to TLP and this is also mentioned in the ietf draft. However, I think they are also complementary in reducing loss recovery delay.

Quoted 5 lines

I've voiced this concern multiple times in ietf tcpm discussion when we discuss this draft: that the idea itself is fine, but we'll need to change Linux RTO algorithm together, not just the timer itself.

Quoted 7 lines
Quoted 4 lines
Quoted 3 lines
Quoted 1 line

I am happy to post some more detailed data if people are interested.

We would love to see the data. The last time we discussed this, you were not

Sure I can gather some data this week for rto-restart.

That would be great! If you will be in London for the IETF perhaps we could also schedule a time to discuss it?

Sure but I should have the data by this weekend. I tweaked your patch about the rto overflow issue.

Quoted 6 lines

I think it's worth the effort to evaluate RTO more thoroughly and preferably use the same re-arming logic for retransmission. In other words, always offset the time elapsed since last sent when re-arming a retransmit timer, and relax the RTO calculation formula but not depending on the inflight. For example, we already use similar logic when handling ICMP dest-unreach. in tcp_ipv{4|6}.c.

I tested your patch on two co-located Google servers in North America. I use all the default upstream TCP features (sack, FACK, tlp, frto, IW10, etc) but vary the rto-restart knob. The latency didn't really change. If anything it's slightly worse but that's within the 1% noise. The retransmission rate are nearly identical. The breakdown of various retranmission stats are the same too and SpuriousRtxHostQueues are negligible. Each experiment has x+-1% millions of samples over a course of 4 days.

*HTTP response latency.
baseline: 673ms, rto-restart:678ms, retrans-rate: 3.6%

HTTP response latency of size <= 3 MSS packets
baseline: 166ms, rto-restart: 171ms, retrans-rate: 3.0%

HTTP response latency is the interval from the first byte of the response is sent out to the last byte is acked.

So it does not hurt nor does it really help in my test.

Quoted 113 lines
Quoted 113 lines
Quoted 14 lines
Quoted 5 lines

able to find it. From the latest discussions in the tcpm group I understood you didn't think it was that big of a problem anymore

rto-restart draft is OK with the context of hypothetical stack implementing only RFC standards. But Linux is (much) more complex:
min-RTO=200ms, TLP, FACK, etc. Changing how the default timer is armed needs to be carefully tested in real world before making rto-restart default. Shortening timer (retry faster) always help short connections but the long connections may suffer the cwnd reset due to spurious timeout, *especially* Linux is already running a very aggressive min-rto.

Quoted 2 lines

btw, the rto -= delta may overflow if the socket was locked when the timer fired previously?

Thanks for pointing that out, I think you might be right.

Best Regards,
Per Hurtig

Quoted 89 lines
Quoted 89 lines

Best Regards,
Per Hurtig

Quoted 83 lines
Quoted 75 lines

The RTO Restart proposal is accepted as a working group item in the IETF TCP Maintenance and Minor Extensions (tcpm) TCP wg and is intended for experimental RFC status.

---

include/net/tcp.h          |    1 +
net/ipv4/sysctl_net_ipv4.c |    7 +++++++
net/ipv4/tcp_input.c       |   11 +++++++++++
3 files changed, 19 insertions(+)
include/net/tcp.h +1 -0
index 56fc366..575e82a 100644
@@ -278,6 +278,7 @@ extern int sysctl_tcp_slow_start_after_idle;
 extern int sysctl_tcp_thin_linear_timeouts;
 extern int sysctl_tcp_thin_dupack;
 extern int sysctl_tcp_early_retrans;
+extern int sysctl_tcp_rto_restart;
 extern int sysctl_tcp_limit_output_bytes;
 extern int sysctl_tcp_challenge_ack_limit;
 extern unsigned int sysctl_tcp_notsent_lowat;
net/ipv4/sysctl_net_ipv4.c +7 -0
index 44eba05..c605f4f 100644
@@ -717,6 +717,13 @@ static struct ctl_table ipv4_table[] = {
               .extra2         = &four,
       },
       {
+               .procname       = "tcp_rto_restart",
+               .data           = &sysctl_tcp_rto_restart,
+               .maxlen         = sizeof(int),
+               .mode           = 0644,
+               .proc_handler   = proc_dointvec,
+       },
+       {
               .procname       = "tcp_min_tso_segs",
               .data           = &sysctl_tcp_min_tso_segs,
               .maxlen         = sizeof(int),
net/ipv4/tcp_input.c +11 -1
index 227cba7..450ee30 100644
@@ -98,6 +98,7 @@ int sysctl_tcp_thin_dupack __read_mostly;
 
 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
 int sysctl_tcp_early_retrans __read_mostly = 3;
+int sysctl_tcp_rto_restart __read_mostly = 1;
 
 #define FLAG_DATA              0x01 /* Incoming frame contained data.          */
 #define FLAG_WIN_UPDATE                0x02 /* Incoming ACK was a window update.       */
@@ -2972,6 +2973,16 @@ void tcp_rearm_rto(struct sock *sk)
                        */
                       if (delta > 0)
                               rto = delta;
+               } else if (icsk->icsk_pending == ICSK_TIME_RETRANS &&
+                          sysctl_tcp_rto_restart &&
+                          sk->sk_send_head == NULL &&
+                          tp->packets_out < 4) {
+                       struct sk_buff *skb = tcp_write_queue_head(sk);
+                       const u32 rto_time_stamp = TCP_SKB_CB(skb)->when;
+                       s32 delta = (s32)(tcp_time_stamp - rto_time_stamp);
+
+                       if (delta > 0)
+                               rto -= delta;
               }
               inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
                                         TCP_RTO_MAX);
--

1.7.9.5

Quoted 4 lines

--
To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

--
To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

author
#7 lore raw
Quoted 84 lines · Yuchung Cheng
Quoted 64 lines · Yuchung Cheng
Quoted 53 lines · Per Hurtig
Quoted 20 lines · Yuchung Cheng
Quoted 19 lines · Per Hurtig

Hi Yuchung,

see inline

Quoted 14 lines · Yuchung Cheng
Quoted 13 lines · Per Hurtig

This patch implements the RTO restart modification described in
http://tools.ietf.org/html/draft-ietf-tcpm-rtorestart-02

RTO Restart's goal is to provide quicker loss recovery for segments lost in the end of a burst/connection. To accomplish this the algorithm adjusts the RTO value on each rearm of the retransmission timer to be exactly RTO ms after the earliest outstanding segment was sent. The offsetting against the earliest outstanding segment is not done by the regular rearm algorithm, which causes RTOs to occur, on average, after RTO+RTT ms.

As a faster timeout is only beneficial in scenarios where fast retransmit/early retransmit cannot be triggered the algorithm will only kick in when there is a small amount of outstanding segments.

Quoted 7 lines
Quoted 7 lines
Quoted 7 lines

(repost in plaine-text, sorry for the duplication)

I am not sure this works well with Linux min-RTO=200ms, and don't feel comfortable this is default on without some real experiments.

I've implemented (a nearly identical version of) rto-restart on Google web servers and found #timeouts increased by ~30%.

How did your version work, did it apply the modified restart for any amount of outstanding segments? If so, I’m not surprised if the #timeouts increased dramatically.

Quoted 17 lines
Quoted 17 lines
Quoted 7 lines

Although the fast
timeout helps really short flows, it has a very negative side-effect:
resetting cwnd to 1. Thus the next HTTP response may start with a small cwnd. In contrast, TCP loss probe has fast timeout (1.5RTT) to do fast recovery to avoid the side-effect. In other words, I am doubtful we need rto-restart with TCP loss probe, but applying RTO-restart on TLP timer may be useful.

As we have discussed, I agree that RTO restart can also be applied to TLP and this is also mentioned in the ietf draft. However, I think they are also complementary in reducing loss recovery delay.

Quoted 5 lines

I've voiced this concern multiple times in ietf tcpm discussion when we discuss this draft: that the idea itself is fine, but we'll need to change Linux RTO algorithm together, not just the timer itself.

Quoted 7 lines
Quoted 4 lines
Quoted 3 lines
Quoted 1 line

I am happy to post some more detailed data if people are interested.

We would love to see the data. The last time we discussed this, you were not

Sure I can gather some data this week for rto-restart.

That would be great! If you will be in London for the IETF perhaps we could also schedule a time to discuss it?

Sure but I should have the data by this weekend. I tweaked your patch about the rto overflow issue.

Quoted 6 lines

I think it's worth the effort to evaluate RTO more thoroughly and preferably use the same re-arming logic for retransmission. In other words, always offset the time elapsed since last sent when re-arming a retransmit timer, and relax the RTO calculation formula but not depending on the inflight. For example, we already use similar logic when handling ICMP dest-unreach. in tcp_ipv{4|6}.c.

I tested your patch on two co-located Google servers in North America. I use all the default upstream TCP features (sack, FACK, tlp, frto, IW10, etc) but vary the rto-restart knob. The latency didn't really change. If anything it's slightly worse but that's within the 1% noise. The retransmission rate are nearly identical. The breakdown of various retranmission stats are the same too and SpuriousRtxHostQueues are negligible. Each experiment has x+-1% millions of samples over a course of 4 days.

*HTTP response latency.
baseline: 673ms, rto-restart:678ms, retrans-rate: 3.6%

Quoted 7 lines

HTTP response latency of size <= 3 MSS packets
baseline: 166ms, rto-restart: 171ms, retrans-rate: 3.0%

HTTP response latency is the interval from the first byte of the response is sent out to the last byte is acked.

So it does not hurt nor does it really help in my test.

Thanks for sharing the data,

Do you know, or is it possible to determine, how the flows that saw loss were affected by the modified restart, in terms of latency? From measurements we’re currently conducting its rather common that the average performance benefit for all flows, including those without loss, is low. However, for flows that do experience loss, there can be significant improvements.

Also, as your setup includes TLP we only have the RTO restart over TLP gain to start with.

Cheers,
Per

Quoted 118 lines
Quoted 103 lines
Quoted 103 lines
Quoted 15 lines
Quoted 6 lines

able to find it. From the latest discussions in the tcpm group I understood you didn't think it was that big of a problem anymore

rto-restart draft is OK with the context of hypothetical stack implementing only RFC standards. But Linux is (much) more complex:
min-RTO=200ms, TLP, FACK, etc. Changing how the default timer is armed needs to be carefully tested in real world before making rto-restart default. Shortening timer (retry faster) always help short connections but the long connections may suffer the cwnd reset due to spurious timeout, *especially* Linux is already running a very aggressive min-rto.

Quoted 2 lines

btw, the rto -= delta may overflow if the socket was locked when the timer fired previously?

Thanks for pointing that out, I think you might be right.

Best Regards,
Per Hurtig

Quoted 80 lines
Quoted 80 lines

Best Regards,
Per Hurtig

Quoted 75 lines
Quoted 75 lines

The RTO Restart proposal is accepted as a working group item in the IETF TCP Maintenance and Minor Extensions (tcpm) TCP wg and is intended for experimental RFC status.

---

include/net/tcp.h          |    1 +
net/ipv4/sysctl_net_ipv4.c |    7 +++++++
net/ipv4/tcp_input.c       |   11 +++++++++++
3 files changed, 19 insertions(+)
include/net/tcp.h +1 -0
index 56fc366..575e82a 100644
@@ -278,6 +278,7 @@ extern int sysctl_tcp_slow_start_after_idle;
 extern int sysctl_tcp_thin_linear_timeouts;
 extern int sysctl_tcp_thin_dupack;
 extern int sysctl_tcp_early_retrans;
+extern int sysctl_tcp_rto_restart;
 extern int sysctl_tcp_limit_output_bytes;
 extern int sysctl_tcp_challenge_ack_limit;
 extern unsigned int sysctl_tcp_notsent_lowat;
net/ipv4/sysctl_net_ipv4.c +7 -0
index 44eba05..c605f4f 100644
@@ -717,6 +717,13 @@ static struct ctl_table ipv4_table[] = {
              .extra2         = &four,
      },
      {
+               .procname       = "tcp_rto_restart",
+               .data           = &sysctl_tcp_rto_restart,
+               .maxlen         = sizeof(int),
+               .mode           = 0644,
+               .proc_handler   = proc_dointvec,
+       },
+       {
              .procname       = "tcp_min_tso_segs",
              .data           = &sysctl_tcp_min_tso_segs,
              .maxlen         = sizeof(int),
net/ipv4/tcp_input.c +11 -1
index 227cba7..450ee30 100644
@@ -98,6 +98,7 @@ int sysctl_tcp_thin_dupack __read_mostly;
 
 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
 int sysctl_tcp_early_retrans __read_mostly = 3;
+int sysctl_tcp_rto_restart __read_mostly = 1;
 
 #define FLAG_DATA              0x01 /* Incoming frame contained data.          */
 #define FLAG_WIN_UPDATE                0x02 /* Incoming ACK was a window update.       */
@@ -2972,6 +2973,16 @@ void tcp_rearm_rto(struct sock *sk)
                       */
                      if (delta > 0)
                              rto = delta;
+               } else if (icsk->icsk_pending == ICSK_TIME_RETRANS &&
+                          sysctl_tcp_rto_restart &&
+                          sk->sk_send_head == NULL &&
+                          tp->packets_out < 4) {
+                       struct sk_buff *skb = tcp_write_queue_head(sk);
+                       const u32 rto_time_stamp = TCP_SKB_CB(skb)->when;
+                       s32 delta = (s32)(tcp_time_stamp - rto_time_stamp);
+
+                       if (delta > 0)
+                               rto -= delta;
              }
              inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
                                        TCP_RTO_MAX);
--

1.7.9.5

Quoted 11 lines
Quoted 11 lines
Quoted 9 lines
Quoted 9 lines
Quoted 8 lines
Quoted 4 lines

--
To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

--
To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

--
To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

maintainer
#8 lore raw
Quoted 2 lines · Per Hurtig

Also, as your setup includes TLP we only have the RTO restart over TLP gain to start with.

My gut says the next thing to try is applying the "RTO restart" idea to the scheduling of the TLP probe (scheduling the TLP probe for write_queue_head_skb->when + tlp_timeout). Doing an RTO earlier can be a very bad thing, due to reseting the cwnd to 1. But doing a TLP earlier does not have much downside, and could have significant latency savings (though maybe Nandita already tried this at some point)...

neal

#9 lore raw
Quoted 11 lines · Neal Cardwell
Quoted 2 lines · Per Hurtig

Also, as your setup includes TLP we only have the RTO restart over TLP gain to start with.

My gut says the next thing to try is applying the "RTO restart" idea to the scheduling of the TLP probe (scheduling the TLP probe for write_queue_head_skb->when + tlp_timeout). Doing an RTO earlier can be a very bad thing, due to reseting the cwnd to 1. But doing a TLP earlier does not have much downside, and could have significant latency savings (though maybe Nandita already tried this at some point)...

+1 for applying RTO restart idea to scheduling a TLP. The thinking with TLP at that time, was to send a probe approximately 1.5RTT after the last "event" (which could be the transmit of a data packet or receipt of an ACK). I did not try scheduling at write_queue_head_skb->when + tlp_timeout, but I think it's worth an experiment (may increase spurious TLPs by some amount which may be OK).

Nandita

Acked-by Tested-by on v1 1/1
#10 lore raw
In reply to Per Hurtig

I’ve applied and tested the patch on VMs with both 32 bit and 64 bit kernel.

Best regards,
Andreas Petlund

Quoted 87 lines · Per Hurtig

This patch implements the RTO restart modification described in
http://tools.ietf.org/html/draft-ietf-tcpm-rtorestart-02

RTO Restart's goal is to provide quicker loss recovery for segments lost in the end of a burst/connection. To accomplish this the algorithm adjusts the RTO value on each rearm of the retransmission timer to be exactly RTO ms after the earliest outstanding segment was sent. The offsetting against the earliest outstanding segment is not done by the regular rearm algorithm, which causes RTOs to occur, on average, after RTO+RTT ms.

As a faster timeout is only beneficial in scenarios where fast retransmit/early retransmit cannot be triggered the algorithm will only kick in when there is a small amount of outstanding segments.

The RTO Restart proposal is accepted as a working group item in the IETF TCP Maintenance and Minor Extensions (tcpm) TCP wg and is intended for experimental RFC status.

---

include/net/tcp.h          |    1 +
net/ipv4/sysctl_net_ipv4.c |    7 +++++++
net/ipv4/tcp_input.c       |   11 +++++++++++
3 files changed, 19 insertions(+)
include/net/tcp.h +1 -0
index 56fc366..575e82a 100644
@@ -278,6 +278,7 @@ extern int sysctl_tcp_slow_start_after_idle;
 extern int sysctl_tcp_thin_linear_timeouts;
 extern int sysctl_tcp_thin_dupack;
 extern int sysctl_tcp_early_retrans;
+extern int sysctl_tcp_rto_restart;
 extern int sysctl_tcp_limit_output_bytes;
 extern int sysctl_tcp_challenge_ack_limit;
 extern unsigned int sysctl_tcp_notsent_lowat;
net/ipv4/sysctl_net_ipv4.c +7 -0
index 44eba05..c605f4f 100644
@@ -717,6 +717,13 @@ static struct ctl_table ipv4_table[] = {
 		.extra2		= &four,
 	},
 	{
+		.procname	= "tcp_rto_restart",
+		.data		= &sysctl_tcp_rto_restart,
+		.maxlen		= sizeof(int),
+		.mode		= 0644,
+		.proc_handler	= proc_dointvec,
+	},
+	{
 		.procname	= "tcp_min_tso_segs",
 		.data		= &sysctl_tcp_min_tso_segs,
 		.maxlen		= sizeof(int),
net/ipv4/tcp_input.c +11 -0
index 227cba7..450ee30 100644
@@ -98,6 +98,7 @@ int sysctl_tcp_thin_dupack __read_mostly;
 
 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
 int sysctl_tcp_early_retrans __read_mostly = 3;
+int sysctl_tcp_rto_restart __read_mostly = 1;
 
 #define FLAG_DATA		0x01 /* Incoming frame contained data.		*/
 #define FLAG_WIN_UPDATE		0x02 /* Incoming ACK was a window update.	*/
@@ -2972,6 +2973,16 @@ void tcp_rearm_rto(struct sock *sk)
 			 */
 			if (delta > 0)
 				rto = delta;
+		} else if (icsk->icsk_pending == ICSK_TIME_RETRANS &&
+			   sysctl_tcp_rto_restart &&
+			   sk->sk_send_head == NULL &&
+			   tp->packets_out < 4) {
+			struct sk_buff *skb = tcp_write_queue_head(sk);
+			const u32 rto_time_stamp = TCP_SKB_CB(skb)->when;
+			s32 delta = (s32)(tcp_time_stamp - rto_time_stamp);
+
+			if (delta > 0)
+				rto -= delta;
 		}
 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
 					  TCP_RTO_MAX);
Signature · 1 line
1.7.9.5
#11 lore raw
Quoted 4 lines · Andreas Petlund

I’ve applied and tested the patch on VMs with both 32 bit and 64 bit kernel.

Sure the patch applies, but what experimental results did you get ?

My concern is that TCP_SKB_CB(skb)->when doesn't account for sojourn time in host queues (qdisc and device).

Note : Missing documentation for this new sysctl.

This patch certainly cannot be applied as is.

Thanks !

#12 lore raw
Quoted 11 lines · Eric Dumazet
Quoted 4 lines · Andreas Petlund

I’ve applied and tested the patch on VMs with both 32 bit and 64 bit kernel.

Sure the patch applies, but what experimental results did you get ?

My concern is that TCP_SKB_CB(skb)->when doesn't account for sojourn time in host queues (qdisc and device).

Thank’s for the input. I didn’t run any specific tests with that in mind. I’ll look into it.

Best regards,
Andreas

author
#13 lore raw
In reply to Eric Dumazet
Quoted 11 lines · Eric Dumazet
Quoted 4 lines · Andreas Petlund

I’ve applied and tested the patch on VMs with both 32 bit and 64 bit kernel.

Sure the patch applies, but what experimental results did you get ?

My concern is that TCP_SKB_CB(skb)->when doesn't account for sojourn time in host queues (qdisc and device).

Correct me if I’m wrong, but I was under the impression that the sojourn was accounted for in the RTT calculations, and therefore rather inappropriate to be accounted for again? If this assumption is wrong, how would you propose that we account for the sojourn time?

Quoted 2 lines

Note : Missing documentation for this new sysctl.

Yes, this we completely missed. Will fix.

Quoted 5 lines

This patch certainly cannot be applied as is.

Thanks !

Thank you,
Per Hurtig

#14 lore raw
Quoted 10 lines · Per Hurtig
Quoted 8 lines · Eric Dumazet
Quoted 4 lines · Andreas Petlund

I’ve applied and tested the patch on VMs with both 32 bit and 64 bit kernel.

Sure the patch applies, but what experimental results did you get ?

Quoted 8 lines
Quoted 3 lines

My concern is that TCP_SKB_CB(skb)->when doesn't account for sojourn time in host queues (qdisc and device).

Correct me if I’m wrong, but I was under the impression that the sojourn was accounted for in the RTT calculations, and therefore rather inappropriate to be accounted for again? If this assumption is wrong, how would you propose that we account for the sojourn time?

This is a generic problem.

Thats why I added LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES snmp counter, so that we can experiment with the idea of not firing RTO while original packet is still in the host qdisc.

       struct sk_buff *skb = tcp_write_queue_head(sk);
       const u32 rto_time_stamp = TCP_SKB_CB(skb)->when;
       s32 delta = (s32)(tcp_time_stamp - rto_time_stamp);

       if (delta > 0)
            rto -= delta;

Can trivially gives rto < 0, resulting in a timer of 120 seconds (TCP_RTO_MAX)

#15 lore raw
In reply to Per Hurtig
Quoted 18 lines · Per Hurtig

This patch implements the RTO restart modification described in
http://tools.ietf.org/html/draft-ietf-tcpm-rtorestart-02

RTO Restart's goal is to provide quicker loss recovery for segments lost in the end of a burst/connection. To accomplish this the algorithm adjusts the RTO value on each rearm of the retransmission timer to be exactly RTO ms after the earliest outstanding segment was sent. The offsetting against the earliest outstanding segment is not done by the regular rearm algorithm, which causes RTOs to occur, on average, after RTO+RTT ms.

As a faster timeout is only beneficial in scenarios where fast retransmit/early retransmit cannot be triggered the algorithm will only kick in when there is a small amount of outstanding segments.

The RTO Restart proposal is accepted as a working group item in the IETF TCP Maintenance and Minor Extensions (tcpm) TCP wg and is intended for experimental RFC status.

Quoted 6 lines
---
 include/net/tcp.h          |    1 +
 net/ipv4/sysctl_net_ipv4.c |    7 +++++++
 net/ipv4/tcp_input.c       |   11 +++++++++++
 3 files changed, 19 insertions(+)

This should also have a documentation update to ip-sysctl.txt. A separate patch for the TCP(7) man page would also be useful if the kernel patch is approved.

-Bill

Quoted 61 lines
include/net/tcp.h +1 -0
index 56fc366..575e82a 100644
@@ -278,6 +278,7 @@ extern int sysctl_tcp_slow_start_after_idle;
 extern int sysctl_tcp_thin_linear_timeouts;
 extern int sysctl_tcp_thin_dupack;
 extern int sysctl_tcp_early_retrans;
+extern int sysctl_tcp_rto_restart;
 extern int sysctl_tcp_limit_output_bytes;
 extern int sysctl_tcp_challenge_ack_limit;
 extern unsigned int sysctl_tcp_notsent_lowat;
net/ipv4/sysctl_net_ipv4.c +7 -0
index 44eba05..c605f4f 100644
@@ -717,6 +717,13 @@ static struct ctl_table ipv4_table[] = {
 		.extra2		= &four,
 	},
 	{
+		.procname	= "tcp_rto_restart",
+		.data		= &sysctl_tcp_rto_restart,
+		.maxlen		= sizeof(int),
+		.mode		= 0644,
+		.proc_handler	= proc_dointvec,
+	},
+	{
 		.procname	= "tcp_min_tso_segs",
 		.data		= &sysctl_tcp_min_tso_segs,
 		.maxlen		= sizeof(int),
net/ipv4/tcp_input.c +11 -0
index 227cba7..450ee30 100644
@@ -98,6 +98,7 @@ int sysctl_tcp_thin_dupack __read_mostly;
 
 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
 int sysctl_tcp_early_retrans __read_mostly = 3;
+int sysctl_tcp_rto_restart __read_mostly = 1;
 
 #define FLAG_DATA		0x01 /* Incoming frame contained data.		*/
 #define FLAG_WIN_UPDATE		0x02 /* Incoming ACK was a window update.	*/
@@ -2972,6 +2973,16 @@ void tcp_rearm_rto(struct sock *sk)
 			 */
 			if (delta > 0)
 				rto = delta;
+		} else if (icsk->icsk_pending == ICSK_TIME_RETRANS &&
+			   sysctl_tcp_rto_restart &&
+			   sk->sk_send_head == NULL &&
+			   tp->packets_out < 4) {
+			struct sk_buff *skb = tcp_write_queue_head(sk);
+			const u32 rto_time_stamp = TCP_SKB_CB(skb)->when;
+			s32 delta = (s32)(tcp_time_stamp - rto_time_stamp);
+
+			if (delta > 0)
+				rto -= delta;
 		}
 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
 					  TCP_RTO_MAX);
Signature · 1 line
1.7.9.5
author
#16 lore raw
Quoted 27 lines · Bill Fink
Quoted 18 lines · Per Hurtig

This patch implements the RTO restart modification described in
http://tools.ietf.org/html/draft-ietf-tcpm-rtorestart-02

RTO Restart's goal is to provide quicker loss recovery for segments lost in the end of a burst/connection. To accomplish this the algorithm adjusts the RTO value on each rearm of the retransmission timer to be exactly RTO ms after the earliest outstanding segment was sent. The offsetting against the earliest outstanding segment is not done by the regular rearm algorithm, which causes RTOs to occur, on average, after RTO+RTT ms.

As a faster timeout is only beneficial in scenarios where fast retransmit/early retransmit cannot be triggered the algorithm will only kick in when there is a small amount of outstanding segments.

The RTO Restart proposal is accepted as a working group item in the IETF TCP Maintenance and Minor Extensions (tcpm) TCP wg and is intended for experimental RFC status.

Quoted 6 lines

---

include/net/tcp.h          |    1 +
net/ipv4/sysctl_net_ipv4.c |    7 +++++++
net/ipv4/tcp_input.c       |   11 +++++++++++
3 files changed, 19 insertions(+)
Quoted 6 lines

This should also have a documentation update to ip-sysctl.txt. A separate patch for the TCP(7) man page would also be useful if the kernel patch is approved.

-Bill

Thanks for pointing that out, we completely forgot it!

Best Regards,
Per Hurtig

Quoted 63 lines
Quoted 63 lines
include/net/tcp.h +1 -0
index 56fc366..575e82a 100644
@@ -278,6 +278,7 @@ extern int sysctl_tcp_slow_start_after_idle;
 extern int sysctl_tcp_thin_linear_timeouts;
 extern int sysctl_tcp_thin_dupack;
 extern int sysctl_tcp_early_retrans;
+extern int sysctl_tcp_rto_restart;
 extern int sysctl_tcp_limit_output_bytes;
 extern int sysctl_tcp_challenge_ack_limit;
 extern unsigned int sysctl_tcp_notsent_lowat;
net/ipv4/sysctl_net_ipv4.c +7 -0
index 44eba05..c605f4f 100644
@@ -717,6 +717,13 @@ static struct ctl_table ipv4_table[] = {
 		.extra2		= &four,
 	},
 	{
+		.procname	= "tcp_rto_restart",
+		.data		= &sysctl_tcp_rto_restart,
+		.maxlen		= sizeof(int),
+		.mode		= 0644,
+		.proc_handler	= proc_dointvec,
+	},
+	{
 		.procname	= "tcp_min_tso_segs",
 		.data		= &sysctl_tcp_min_tso_segs,
 		.maxlen		= sizeof(int),
net/ipv4/tcp_input.c +11 -0
index 227cba7..450ee30 100644
@@ -98,6 +98,7 @@ int sysctl_tcp_thin_dupack __read_mostly;
 
 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
 int sysctl_tcp_early_retrans __read_mostly = 3;
+int sysctl_tcp_rto_restart __read_mostly = 1;
 
 #define FLAG_DATA		0x01 /* Incoming frame contained data.		*/
 #define FLAG_WIN_UPDATE		0x02 /* Incoming ACK was a window update.	*/
@@ -2972,6 +2973,16 @@ void tcp_rearm_rto(struct sock *sk)
 			 */
 			if (delta > 0)
 				rto = delta;
+		} else if (icsk->icsk_pending == ICSK_TIME_RETRANS &&
+			   sysctl_tcp_rto_restart &&
+			   sk->sk_send_head == NULL &&
+			   tp->packets_out < 4) {
+			struct sk_buff *skb = tcp_write_queue_head(sk);
+			const u32 rto_time_stamp = TCP_SKB_CB(skb)->when;
+			s32 delta = (s32)(tcp_time_stamp - rto_time_stamp);
+
+			if (delta > 0)
+				rto -= delta;
 		}
 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
 					  TCP_RTO_MAX);
Signature · 1 line
1.7.9.5