net-tcp: Disable TCP ssthresh metrics cache by default
authorKevin(Yudong) Yang <yyd@google.com>
Mon, 9 Dec 2019 19:19:59 +0000 (14:19 -0500)
committerDavid S. Miller <davem@davemloft.net>
Tue, 10 Dec 2019 04:17:48 +0000 (20:17 -0800)
commit65e6d90168f3593df0ae598502bcbf20d78ff0fb
treef8a7670488d31167eb0fdce6033859caee7de056
parent4e7696d90b51a1a73ce0e8174f3aff58b914619c
net-tcp: Disable TCP ssthresh metrics cache by default

This patch introduces a sysctl knob "net.ipv4.tcp_no_ssthresh_metrics_save"
that disables TCP ssthresh metrics cache by default. Other parts of TCP
metrics cache, e.g. rtt, cwnd, remain unchanged.

As modern networks becoming more and more dynamic, TCP metrics cache
today often causes more harm than benefits. For example, the same IP
address is often shared by different subscribers behind NAT in residential
networks. Even if the IP address is not shared by different users,
caching the slow-start threshold of a previous short flow using loss-based
congestion control (e.g. cubic) often causes the future longer flows of
the same network path to exit slow-start prematurely with abysmal
throughput.

Caching ssthresh is very risky and can lead to terrible performance.
Therefore it makes sense to make disabling ssthresh caching by
default and opt-in for specific networks by the administrators.
This practice also has worked well for several years of deployment with
CUBIC congestion control at Google.

Acked-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Kevin(Yudong) Yang <yyd@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Documentation/networking/ip-sysctl.txt
include/net/netns/ipv4.h
net/ipv4/sysctl_net_ipv4.c
net/ipv4/tcp_ipv4.c
net/ipv4/tcp_metrics.c