Skip to contentSkip to conversation

TCP Limited slow start

RFC In review Stephen Hemminger 2 participants 4 messages Jun 2, 2006 – Jun 5, 2006

Subsystems (5)
lore
Share & tools
lore
#0 lore raw

Has anyone done an implementation of RFC3742 for Linux? It looks interesting, but would need some integration with current ABC code.

There was some evidence of a version in old Web100 code, but it's gone now. Was it deemed a mistake?

#1 lore raw

RFC 1/1 TCP limited slow start

Rolled my sleeve's up and gave this a try...

This is a implementation of Sally Floyd's Limited Slow Start for Large Congestion Windows.

Summary from RFC:
Limited Slow-Start introduces a parameter, "max_ssthresh", and modifies the slow-start mechanism for values of the congestion window where "cwnd" is greater than "max_ssthresh". That is, during Slow- Start, when

cwnd <= max_ssthresh,

cwnd is increased by one MSS (MAXIMUM SEGMENT SIZE) for every arriving ACK (acknowledgement) during slow-start, as is always the case. During Limited Slow-Start, when

max_ssthresh < cwnd <= ssthresh,

the invariant is maintained so that the congestion window is increased during slow-start by at most max_ssthresh/2 MSS per round- trip time. This is done as follows:

      For each arriving ACK in slow-start:
        If (cwnd <= max_ssthresh)
           cwnd += MSS;
        else
           K = int(cwnd/(0.5 max_ssthresh));
           cwnd += int(MSS/K);

Thus, during Limited Slow-Start the window is increased by 1/K MSS for each arriving ACK, for K = int(cwnd/(0.5 max_ssthresh)), instead of by 1 MSS as in standard slow-start [RFC2581].

---

 Documentation/networking/ip-sysctl.txt |    8 +++++-
 include/linux/sysctl.h                 |    1 +
 include/net/tcp.h                      |    1 +
 net/ipv4/sysctl_net_ipv4.c             |    8 ++++++
 net/ipv4/tcp_cong.c                    |   46 ++++++++++++++++++++------------
 net/ipv4/tcp_input.c                   |    1 +
 6 files changed, 47 insertions(+), 18 deletions(-)

0884f45c9f21c50dd9117b2fc02bf5436be3c3bf

Documentation/networking/ip-sysctl.txt +7 -1
index f12007b..9869298 100644
@@ -103,9 +103,15 @@ TCP variables: 
 
 tcp_abc - INTEGER
 	Controls Appropriate Byte Count defined in RFC3465. If set to
-	0 then does congestion avoid once per ack. 1 is conservative
+	0 then does congestion avoid once per ack. 1 (default) is conservative
 	value, and 2 is more agressive.
 
+tcp_limited_ssthresh - INTEGER
+	Controls the increase of the congestion window during slow start as
+	defined in RFC3742. The purpose is to slow the growth of the congestion
+	window on high delay networks where agressive growth can cause losses
+	of 1000's of packets. Default is 100 packets.
+
 tcp_syn_retries - INTEGER
 	Number of times initial SYNs for an active TCP connection attempt
 	will be retransmitted. Should not be higher than 255. Default value
include/linux/sysctl.h +1 -0
index 76eaeff..a455165 100644
@@ -403,6 +403,7 @@ enum
  	NET_TCP_MTU_PROBING=113,
 	NET_TCP_BASE_MSS=114,
 	NET_IPV4_TCP_WORKAROUND_SIGNED_WINDOWS=115,
+	NET_TCP_LIMITED_SSTHRESH=116,
 };
 
 enum {
include/net/tcp.h +1 -0
index 575636f..3a14861 100644
@@ -225,6 +225,7 @@ extern int sysctl_tcp_abc;
 extern int sysctl_tcp_mtu_probing;
 extern int sysctl_tcp_base_mss;
 extern int sysctl_tcp_workaround_signed_windows;
+extern int sysctl_tcp_limited_ssthresh;
 
 extern atomic_t tcp_memory_allocated;
 extern atomic_t tcp_sockets_allocated;
net/ipv4/sysctl_net_ipv4.c +8 -0
index 6b6c3ad..d1358d3 100644
@@ -688,6 +688,14 @@ #endif
 		.mode		= 0644,
 		.proc_handler	= &proc_dointvec
 	},
+	{
+		.ctl_name	= NET_TCP_LIMITED_SSTHRESH,
+		.procname	= "tcp_max_ssthresh",
+		.data		= &sysctl_tcp_limited_ssthresh,
+		.maxlen		= sizeof(int),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec,
+	},
 	{ .ctl_name = 0 }
 };
 
net/ipv4/tcp_cong.c +29 -17

54 lines not on this page · raw message

net/ipv4/tcp_input.c +1 -0
index 718d0f2..80dd5e4 100644
@@ -90,6 +90,7 @@ int sysctl_tcp_nometrics_save;
 
 int sysctl_tcp_moderate_rcvbuf = 1;
 int sysctl_tcp_abc = 1;
+int sysctl_tcp_limited_ssthresh = 100;
 
 #define FLAG_DATA		0x01 /* Incoming frame contained data.		*/
 #define FLAG_WIN_UPDATE		0x02 /* Incoming ACK was a window update.	*/
Signature · 1 line
1.3.3
#2 lore raw

Re: [RFC] TCP limited slow start

Quoted 4 lines · Stephen Hemminger

Rolled my sleeve's up and gave this a try...

This is a implementation of Sally Floyd's Limited Slow Start for Large Congestion Windows.

Limited slow start is useful as a work-around for bottleneck queues that are inappropriately short. I don't think it's good to run it all the time by default (with a max_ssthresh < infinity), because it slows down flows on healthy paths, and introduces another non-scalable parameter to TCP.

I see it as potentially useful as a per-route parameter, where you set it deliberately to work around some known problematic path. A sysctl with a default value of infinity might be okay as well.

Practically speaking, we've had this in the Web100 patch for a long time (and still do, look for WAD_MaxSsthresh), but I've never found it all that useful. If the bottleneck queue is too short, you usually end up getting screwed other ways too.

-John

#3 lore raw

Re: [RFC] TCP limited slow start

Quoted 16 lines · On Sat, 03 Jun 2006 12:46:57 -0400 John Heffner <[email protected]> wrote:
Quoted 4 lines · Stephen Hemminger

Rolled my sleeve's up and gave this a try...

This is a implementation of Sally Floyd's Limited Slow Start for Large Congestion Windows.

Limited slow start is useful as a work-around for bottleneck queues that are inappropriately short. I don't think it's good to run it all the time by default (with a max_ssthresh < infinity), because it slows down flows on healthy paths, and introduces another non-scalable parameter to TCP.

I see it as potentially useful as a per-route parameter, where you set it deliberately to work around some known problematic path. A sysctl with a default value of infinity might be okay as well.

Quoted 6 lines

Practically speaking, we've had this in the Web100 patch for a long time (and still do, look for WAD_MaxSsthresh), but I've never found it all that useful. If the bottleneck queue is too short, you usually end up getting screwed other ways too.

-John

I moved it off to tcp_highspeed.c only. That is seems appropriate because that is where you put the related RFC.