The linux packet socket can handle GSO packets, so we can segment packets to
64K instead of the MTU which is usually 1500.
Here are numbers for the nginx-1m test:
runsc: 579330.01 [Kbytes/sec] received
runsc-gso: 1794121.66 [Kbytes/sec] received
runc: 2122139.06 [Kbytes/sec] received
and for tcp_benchmark:
$ tcp_benchmark --duration 15 --ideal
[ 4] 0.0-15.0 sec 86647 MBytes 48456 Mbits/sec
$ tcp_benchmark --client --duration 15 --ideal
[ 4] 0.0-15.0 sec 2173 MBytes 1214 Mbits/sec
$ tcp_benchmark --client --duration 15 --ideal --gso 65536
[ 4] 0.0-15.0 sec 19357 MBytes 10825 Mbits/sec
PiperOrigin-RevId: 240809103
Change-Id: I2637f104db28b5d4c64e1e766c610162a195775a
This change does not make use of SACK information but adds support to track
SACK information and store it in the endpoint.
The actual SACK based recovery will be in a separate CL.
Part of commits to add RFC 6675 support to Netstack.
PiperOrigin-RevId: 235612264
Change-Id: I261f94844d7bad5abda803152ce6cc6125a467ff
RFC7323 recommends that if the timestamp option was negotiated
then all packets should carry a TCP Timestamp and any packets that
do not should be dropped.
Netstack implemented this behaviour. Linux OTOH does not and will
accept such packets. This change makes Netstack behaviour compatible
with Linux.
Also now that we allow such packets, we do need to update RTO calculations
based on these packets even if timestamp option is enabled.
PiperOrigin-RevId: 233432268
Change-Id: I9f4742ae6b63930ac3b5e37d8c238761e6a4b29f
Previously, TCP_NODELAY was always enabled and we would lie about it being
configurable. TCP_NODELAY is now disabled by default (to match Linux) in the
socket layer so that non-gVisor users don't automatically start using this
questionable optimization.
PiperOrigin-RevId: 221368472
Change-Id: Ib0240f66d94455081f4e0ca94f09d9338b2c1356
From RFC7323#Section-4
The [RFC6298] RTT estimator has weighting factors, alpha and beta, based on an
implicit assumption that at most one RTTM will be sampled per RTT. When
multiple RTTMs per RTT are available to update the RTT estimator, an
implementation SHOULD try to adhere to the spirit of the history specified in
[RFC6298]. An implementation suggestion is detailed in Appendix G.
From RFC7323#appendix-G
Appendix G. RTO Calculation Modification
Taking multiple RTT samples per window would shorten the history calculated
by the RTO mechanism in [RFC6298], and the below algorithm aims to maintain a
similar history as originally intended by [RFC6298].
It is roughly known how many samples a congestion window worth of data will
yield, not accounting for ACK compression, and ACK losses. Such events will
result in more history of the path being reflected in the final value for
RTO, and are uncritical. This modification will ensure that a similar amount
of time is taken into account for the RTO estimation, regardless of how many
samples are taken per window:
ExpectedSamples = ceiling(FlightSize / (SMSS * 2))
alpha' = alpha / ExpectedSamples
beta' = beta / ExpectedSamples
Note that the factor 2 in ExpectedSamples is due to "Delayed ACKs".
Instead of using alpha and beta in the algorithm of [RFC6298], use alpha' and
beta' instead:
RTTVAR <- (1 - beta') * RTTVAR + beta' * |SRTT - R'|
SRTT <- (1 - alpha') * SRTT + alpha' * R'
(for each sample R')
PiperOrigin-RevId: 213644795
Change-Id: I52278b703540408938a8edb8c38be97b37f4a10e
Makes it possible to avoid copying or allocating in cases where DeliverNetworkPacket (rx)
needs to turn around and call WritePacket (tx) with its VectorisedView.
Also removes the restriction on having VectorisedViews with multiple views in the write path.
PiperOrigin-RevId: 211728717
Change-Id: Ie03a65ecb4e28bd15ebdb9c69f05eced18fdfcff
This CL implements CUBIC as described in https://tools.ietf.org/html/rfc8312.
PiperOrigin-RevId: 207353142
Change-Id: I329cbf3277f91127e99e488f07d906f6779c6603
This CL also puts the congestion control logic behind an
interface so that we can easily swap it out for say CUBIC
in the future.
PiperOrigin-RevId: 205732848
Change-Id: I891cdfd17d4d126b658b5faa0c6bd6083187944b
There is a subtle bug where during cleanup with unread data a FIN can
be converted to a RST, at that point the entire connection should be
aborted as we're not expecting any ACKs to the RST.
PiperOrigin-RevId: 202691271
Change-Id: Idae70800208ca26e07a379bc6b2b8090805d0a22
Today when we transmit a RST it's happening during the time-wait
flow. Because a FIN is allowed to advance the acceptable ACK window
we're incorrectly doing that for a RST.
PiperOrigin-RevId: 197637565
Change-Id: I080190b06bd0225326cd68c1fbf37bd3fdbd414e
Andrei Vagin
Tamir Duberstein
Bhasker Hariharan
Ian Gudger
Bert Muthalaly
Zhaozhong Ni
Michael Pratt
Nicolas Lacasse
Brian Geffon
Cyrille Hemidy
Googler