[RFC] cubic: backoff after slow start

[RFC] cubic: backoff after slow start

[Stephen Hemminger]

CUBIC takes several unnecessary iterations to converge out of slow start. This
is most noticable over a link where the bottleneck queue size is much larger than BDP,
and the sender has to "fill the pipe" in slow start before the first loss. Typical
consumer broadband links seem to have large (up to 2secs) of queue that needs
to get filled before the first loss.

A possible fix is to use a beta of .5 (same as original TCP) when leaving
slow start.  Originally, the Linux version didn't do slow start so it probably
never was observed.

--- a/net/ipv4/tcp_cubic.c	2007-08-02 12:16:22.000000000 +0100
+++ b/net/ipv4/tcp_cubic.c	2007-08-03 15:57:12.000000000 +0100
@@ -289,7 +289,11 @@ static u32 bictcp_recalc_ssthresh(struct
 
 	ca->loss_cwnd = tp->snd_cwnd;
 
-	return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
+	/* Initial backoff when leaving slow start */
+	if (tp->snd_ssthresh == 0x7fffffff)
+		return max(tp->snd_cwnd >> 1U, 2U);
+	else
+		return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
 }
 
 static u32 bictcp_undo_cwnd(struct sock *sk)

Re: [RFC] cubic: backoff after slow start

[Injong Rhee]

Hi Stephen,
We have been working on slow start and we have a nice solution for this. We 
will send you a patch and test results.
Thanks
Injong


----- Original Message ----- 
From: "Stephen Hemminger" <shemminger@linux-foundation.org>
To: "Injong Rhee" <rhee@eos.ncsu.edu>; "Sangtae Ha" <sha2@ncsu.edu>
Cc: <netdev@vger.kernel.org>
Sent: Tuesday, August 07, 2007 2:37 PM
Subject: [RFC] cubic: backoff after slow start


> CUBIC takes several unnecessary iterations to converge out of slow start. 
> This
> is most noticable over a link where the bottleneck queue size is much 
> larger than BDP,
> and the sender has to "fill the pipe" in slow start before the first loss. 
> Typical
> consumer broadband links seem to have large (up to 2secs) of queue that 
> needs
> to get filled before the first loss.
>
> A possible fix is to use a beta of .5 (same as original TCP) when leaving
> slow start.  Originally, the Linux version didn't do slow start so it 
> probably
> never was observed.
>
> --- a/net/ipv4/tcp_cubic.c 2007-08-02 12:16:22.000000000 +0100
> +++ b/net/ipv4/tcp_cubic.c 2007-08-03 15:57:12.000000000 +0100
> @@ -289,7 +289,11 @@ static u32 bictcp_recalc_ssthresh(struct
>
>  ca->loss_cwnd = tp->snd_cwnd;
>
> - return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
> + /* Initial backoff when leaving slow start */
> + if (tp->snd_ssthresh == 0x7fffffff)
> + return max(tp->snd_cwnd >> 1U, 2U);
> + else
> + return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
> }
>
> static u32 bictcp_undo_cwnd(struct sock *sk)
>