Fix BBR so that upon notification of a loss recovery undo BBR resets
long-term bandwidth sampling.
Under high reordering, reordering events can be interpreted as loss.
If the reordering and spurious loss estimates are high enough, this
can cause BBR to spuriously estimate that we are seeing loss rates
high enough to trigger long-term bandwidth estimation. To avoid that
problem, this commit resets long-term bandwidth sampling on loss
recovery undo events.
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fix BBR so that upon notification of a loss recovery undo BBR resets
the full pipe detection (STARTUP exit) state machine.
Under high reordering, reordering events can be interpreted as loss.
If the reordering and spurious loss estimates are high enough, this
could previously cause BBR to spuriously estimate that the pipe is
full.
Since spurious loss recovery means that our overall sending will have
slowed down spuriously, this commit gives a flow more time to probe
robustly for bandwidth and decide the pipe is really full.
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This commit records the "full bw reached" decision in a new
full_bw_reached bit. This is a pure refactor that does not change the
current behavior, but enables subsequent fixes and improvements.
In particular, this enables simple and clean fixes because the full_bw
and full_bw_cnt can be unconditionally zeroed without worrying about
forgetting that we estimated we filled the pipe in Startup. And it
enables future improvements because multiple code paths can be used
for estimating that we filled the pipe in Startup; any new code paths
only need to set this bit when they think the pipe is full.
Note that this fix intentionally reduces the width of the full_bw_cnt
counter, since we have never used the most significant bit.
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fixes the following behavior: for connections that had no RTT sample
at the time of initializing congestion control, BBR was initializing
the pacing rate to a high nominal rate (based an a guess of RTT=1ms,
in case this is LAN traffic). Then BBR never adjusted the pacing rate
downward upon obtaining an actual RTT sample, if the connection never
filled the pipe (e.g. all sends were small app-limited writes()).
This fix adjusts the pacing rate upon obtaining the first RTT sample.
Fixes: 0f8782ea14 ("tcp_bbr: add BBR congestion control")
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fix a corner case noticed by Eric Dumazet, where BBR's setting
sk->sk_pacing_rate to 0 during initialization could theoretically
cause packets in the sending host to hang if there were packets "in
flight" in the pacing infrastructure at the time the BBR congestion
control state is initialized. This could occur if the pacing
infrastructure happened to race with bbr_init() in a way such that the
pacer read the 0 rather than the immediately following non-zero pacing
rate.
Fixes: 0f8782ea14 ("tcp_bbr: add BBR congestion control")
Reported-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Introduce a helper to initialize the BBR pacing rate unconditionally,
based on the current cwnd and RTT estimate. This is a pure refactor,
but is needed for two following fixes.
Fixes: 0f8782ea14 ("tcp_bbr: add BBR congestion control")
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Introduce a helper to convert a BBR bandwidth and gain factor to a
pacing rate in bytes per second. This is a pure refactor, but is
needed for two following fixes.
Fixes: 0f8782ea14 ("tcp_bbr: add BBR congestion control")
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In bbr_set_pacing_rate(), which decides whether to cut the pacing
rate, there was some code that considered exiting STARTUP to be
equivalent to the notion of filling the pipe (i.e.,
bbr_full_bw_reached()). Specifically, as the code was structured,
exiting STARTUP and going into PROBE_RTT could cause us to cut the
pacing rate down to something silly and low, based on whatever
bandwidth samples we've had so far, when it's possible that all of
them have been small app-limited bandwidth samples that are not
representative of the bandwidth available in the path. (The code was
correct at the time it was written, but the state machine changed
without this spot being adjusted correspondingly.)
Fixes: 0f8782ea14 ("tcp_bbr: add BBR congestion control")
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
TCP Timestamps option is defined in RFC 7323
Traditionally on linux, it has been tied to the internal
'jiffies' variable, because it had been a cheap and good enough
generator.
For TCP flows on the Internet, 1 ms resolution would be much better
than 4ms or 10ms (HZ=250 or HZ=100 respectively)
For TCP flows in the DC, Google has used usec resolution for more
than two years with great success [1]
Receive size autotuning (DRS) is indeed more precise and converges
faster to optimal window size.
This patch converts tp->tcp_mstamp to a plain u64 value storing
a 1 usec TCP clock.
This choice will allow us to upstream the 1 usec TS option as
discussed in IETF 97.
[1] https://www.ietf.org/proceedings/97/slides/slides-97-tcpm-tcp-options-for-low-latency-00.pdf
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use tcp_jiffies32 instead of tcp_time_stamp, since
tcp_time_stamp will soon be only used for TCP TS option.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
BBR congestion control depends on pacing, and pacing is
currently handled by sch_fq packet scheduler for performance reasons,
and also because implemening pacing with FQ was convenient to truly
avoid bursts.
However there are many cases where this packet scheduler constraint
is not practical.
- Many linux hosts are not focusing on handling thousands of TCP
flows in the most efficient way.
- Some routers use fq_codel or other AQM, but still would like
to use BBR for the few TCP flows they initiate/terminate.
This patch implements an automatic fallback to internal pacing.
Pacing is requested either by BBR or use of SO_MAX_PACING_RATE option.
If sch_fq happens to be in the egress path, pacing is delegated to
the qdisc, otherwise pacing is done by TCP itself.
One advantage of pacing from TCP stack is to get more precise rtt
estimations, and less work done from TX completion, since TCP Small
queue limits are not generally hit. Setups with single TX queue but
many cpus might even benefit from this.
Note that unlike sch_fq, we do not take into account header sizes.
Taking care of these headers would add additional complexity for
no practical differences in behavior.
Some performance numbers using 800 TCP_STREAM flows rate limited to
~48 Mbit per second on 40Gbit NIC.
If MQ+pfifo_fast is used on the NIC :
$ sar -n DEV 1 5 | grep eth
14:48:44 eth0 725743.00 2932134.00 46776.76 4335184.68 0.00 0.00 1.00
14:48:45 eth0 725349.00 2932112.00 46751.86 4335158.90 0.00 0.00 0.00
14:48:46 eth0 725101.00 2931153.00 46735.07 4333748.63 0.00 0.00 0.00
14:48:47 eth0 725099.00 2931161.00 46735.11 4333760.44 0.00 0.00 1.00
14:48:48 eth0 725160.00 2931731.00 46738.88 4334606.07 0.00 0.00 0.00
Average: eth0 725290.40 2931658.20 46747.54 4334491.74 0.00 0.00 0.40
$ vmstat 1 5
procs -----------memory---------- ---swap-- -----io---- -system-- ------cpu-----
r b swpd free buff cache si so bi bo in cs us sy id wa st
4 0 0 259825920 45644 2708324 0 0 21 2 247 98 0 0 100 0 0
4 0 0 259823744 45644 2708356 0 0 0 0 2400825 159843 0 19 81 0 0
0 0 0 259824208 45644 2708072 0 0 0 0 2407351 159929 0 19 81 0 0
1 0 0 259824592 45644 2708128 0 0 0 0 2405183 160386 0 19 80 0 0
1 0 0 259824272 45644 2707868 0 0 0 32 2396361 158037 0 19 81 0 0
Now use MQ+FQ :
lpaa23:~# echo fq >/proc/sys/net/core/default_qdisc
lpaa23:~# tc qdisc replace dev eth0 root mq
$ sar -n DEV 1 5 | grep eth
14:49:57 eth0 678614.00 2727930.00 43739.13 4033279.14 0.00 0.00 0.00
14:49:58 eth0 677620.00 2723971.00 43674.69 4027429.62 0.00 0.00 1.00
14:49:59 eth0 676396.00 2719050.00 43596.83 4020125.02 0.00 0.00 0.00
14:50:00 eth0 675197.00 2714173.00 43518.62 4012938.90 0.00 0.00 1.00
14:50:01 eth0 676388.00 2719063.00 43595.47 4020171.64 0.00 0.00 0.00
Average: eth0 676843.00 2720837.40 43624.95 4022788.86 0.00 0.00 0.40
$ vmstat 1 5
procs -----------memory---------- ---swap-- -----io---- -system-- ------cpu-----
r b swpd free buff cache si so bi bo in cs us sy id wa st
2 0 0 259832240 46008 2710912 0 0 21 2 223 192 0 1 99 0 0
1 0 0 259832896 46008 2710744 0 0 0 0 1702206 198078 0 17 82 0 0
0 0 0 259830272 46008 2710596 0 0 0 0 1696340 197756 1 17 83 0 0
4 0 0 259829168 46024 2710584 0 0 16 0 1688472 197158 1 17 82 0 0
3 0 0 259830224 46024 2710408 0 0 0 0 1692450 197212 0 18 82 0 0
As expected, number of interrupts per second is very different.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Van Jacobson <vanj@google.com>
Cc: Jerry Chu <hkchu@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Document the possible state transitions for a BBR flow, and also add a
prose summary of the state machine, covering the life of a typical BBR
flow.
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This commit implements a new TCP congestion control algorithm: BBR
(Bottleneck Bandwidth and RTT). A detailed description of BBR will be
published in ACM Queue, Vol. 14 No. 5, September-October 2016, as
"BBR: Congestion-Based Congestion Control".
BBR has significantly increased throughput and reduced latency for
connections on Google's internal backbone networks and google.com and
YouTube Web servers.
BBR requires only changes on the sender side, not in the network or
the receiver side. Thus it can be incrementally deployed on today's
Internet, or in datacenters.
The Internet has predominantly used loss-based congestion control
(largely Reno or CUBIC) since the 1980s, relying on packet loss as the
signal to slow down. While this worked well for many years, loss-based
congestion control is unfortunately out-dated in today's networks. On
today's Internet, loss-based congestion control causes the infamous
bufferbloat problem, often causing seconds of needless queuing delay,
since it fills the bloated buffers in many last-mile links. On today's
high-speed long-haul links using commodity switches with shallow
buffers, loss-based congestion control has abysmal throughput because
it over-reacts to losses caused by transient traffic bursts.
In 1981 Kleinrock and Gale showed that the optimal operating point for
a network maximizes delivered bandwidth while minimizing delay and
loss, not only for single connections but for the network as a
whole. Finding that optimal operating point has been elusive, since
any single network measurement is ambiguous: network measurements are
the result of both bandwidth and propagation delay, and those two
cannot be measured simultaneously.
While it is impossible to disambiguate any single bandwidth or RTT
measurement, a connection's behavior over time tells a clearer
story. BBR uses a measurement strategy designed to resolve this
ambiguity. It combines these measurements with a robust servo loop
using recent control systems advances to implement a distributed
congestion control algorithm that reacts to actual congestion, not
packet loss or transient queue delay, and is designed to converge with
high probability to a point near the optimal operating point.
In a nutshell, BBR creates an explicit model of the network pipe by
sequentially probing the bottleneck bandwidth and RTT. On the arrival
of each ACK, BBR derives the current delivery rate of the last round
trip, and feeds it through a windowed max-filter to estimate the
bottleneck bandwidth. Conversely it uses a windowed min-filter to
estimate the round trip propagation delay. The max-filtered bandwidth
and min-filtered RTT estimates form BBR's model of the network pipe.
Using its model, BBR sets control parameters to govern sending
behavior. The primary control is the pacing rate: BBR applies a gain
multiplier to transmit faster or slower than the observed bottleneck
bandwidth. The conventional congestion window (cwnd) is now the
secondary control; the cwnd is set to a small multiple of the
estimated BDP (bandwidth-delay product) in order to allow full
utilization and bandwidth probing while bounding the potential amount
of queue at the bottleneck.
When a BBR connection starts, it enters STARTUP mode and applies a
high gain to perform an exponential search to quickly probe the
bottleneck bandwidth (doubling its sending rate each round trip, like
slow start). However, instead of continuing until it fills up the
buffer (i.e. a loss), or until delay or ACK spacing reaches some
threshold (like Hystart), it uses its model of the pipe to estimate
when that pipe is full: it estimates the pipe is full when it notices
the estimated bandwidth has stopped growing. At that point it exits
STARTUP and enters DRAIN mode, where it reduces its pacing rate to
drain the queue it estimates it has created.
Then BBR enters steady state. In steady state, PROBE_BW mode cycles
between first pacing faster to probe for more bandwidth, then pacing
slower to drain any queue that created if no more bandwidth was
available, and then cruising at the estimated bandwidth to utilize the
pipe without creating excess queue. Occasionally, on an as-needed
basis, it sends significantly slower to probe for RTT (PROBE_RTT
mode).
BBR has been fully deployed on Google's wide-area backbone networks
and we're experimenting with BBR on Google.com and YouTube on a global
scale. Replacing CUBIC with BBR has resulted in significant
improvements in network latency and application (RPC, browser, and
video) metrics. For more details please refer to our upcoming ACM
Queue publication.
Example performance results, to illustrate the difference between BBR
and CUBIC:
Resilience to random loss (e.g. from shallow buffers):
Consider a netperf TCP_STREAM test lasting 30 secs on an emulated
path with a 10Gbps bottleneck, 100ms RTT, and 1% packet loss
rate. CUBIC gets 3.27 Mbps, and BBR gets 9150 Mbps (2798x higher).
Low latency with the bloated buffers common in today's last-mile links:
Consider a netperf TCP_STREAM test lasting 120 secs on an emulated
path with a 10Mbps bottleneck, 40ms RTT, and 1000-packet bottleneck
buffer. Both fully utilize the bottleneck bandwidth, but BBR
achieves this with a median RTT 25x lower (43 ms instead of 1.09
secs).
Our long-term goal is to improve the congestion control algorithms
used on the Internet. We are hopeful that BBR can help advance the
efforts toward this goal, and motivate the community to do further
research.
Test results, performance evaluations, feedback, and BBR-related
discussions are very welcome in the public e-mail list for BBR:
https://groups.google.com/forum/#!forum/bbr-dev
NOTE: BBR *must* be used with the fq qdisc ("man tc-fq") with pacing
enabled, since pacing is integral to the BBR design and
implementation. BBR without pacing would not function properly, and
may incur unnecessary high packet loss rates.
Signed-off-by: Van Jacobson <vanj@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Nandita Dukkipati <nanditad@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Neal Cardwell