Re: [PATCH net-next v3] tcp: use RFC6298 compliant TCP RTO calculation

[Daniel Metz]

  BD  R Y
--
mean TCPRecovLat  3s  -7%   +39% +38%
mean TCPRecovLat252s +1%   -11%  -11%


This is indeed very interesting and somewhat unexpected. Do you have any 
clue why Y is as bad as R and so much worse than B? By my understanding 
I would have expected Y to be similar to B. At least tests on the Mean 
Response Waiting Time of sender limited flows show hardly any difference 
to B (as expected).


Also, is a potential longer time in TCPRecovLat such a bad thing 
considering your information on HTTP response performance?

Re: [PATCH net-next v3] tcp: use RFC6298 compliant TCP RTO calculation

[Yuchung Cheng]

On Tue, Jun 21, 2016 at 10:53 PM, Yuchung Cheng  wrote:
>
> On Fri, Jun 17, 2016 at 11:56 AM, Yuchung Cheng  wrote:
> >
> > On Fri, Jun 17, 2016 at 11:32 AM, David Miller  wrote:
> > >
> > > From: Daniel Metz 
> > > Date: Wed, 15 Jun 2016 20:00:03 +0200
> > >
> > > > This patch adjusts Linux RTO calculation to be RFC6298 Standard
> > > > compliant. MinRTO is no longer added to the computed RTO, RTO damping
> > > > and overestimation are decreased.
> > >  ...
> > >
> > > Yuchung, I assume I am waiting for you to do the testing you said
> > > you would do for this patch, right?
> > Yes I spent the last two days resolving some unrelated glitches to
> > start my testing on Web servers. I should be able to get some results
> > over the weekend.
> >
> > I will test
> > 0) current Linux
> > 1) this patch
> > 2) RFC6298 with min_RTO=1sec
> > 3) RFC6298 with minimum RTTVAR of 200ms (so it is more like current
> > Linux style of min RTO which only applies to RTTVAR)
> >
> > and collect the TCP latency (how long to send an HTTP response) and
> > (spurious) timeout & retransmission stats.
> >
> Thanks for the patience. I've collected data from some Google Web
> servers. They serve both a mix of US and SouthAm users using
> HTTP1 and HTTP2. The traffic is Web browsing (e.g., search, maps,
> gmails, etc but not Youtube videos). The mean RTT is about 100ms.
>
> The user connections were split into 4 groups of different TCP RTO
> configs. Each group has many millions of connections but the
> size variation among groups is well under 1%.
>
> B: baseline Linux
> D: this patch
> R: change RTTYAR averaging as in D, but bound RTO to 1sec per RFC6298
> Y: change RTTVAR averaging as in D, but bound RTTVAR to 200ms instead (like B)
>
> For mean TCP latency of HTTP responses (first byte sent to last byte
> acked), B < R < Y < D. But the differences are so insignificant (<1%).
> The median, 95pctl, and 99pctl has similar indifference. In summary
> there's hardly visible impact on latency. I also look at only response
> less than 4KB but do not see a different picture.
>
> The main difference is the retransmission rate where R =~ Y < B =~D.
> R and Y are ~20% lower than B and D. Parsing the SNMP stats reveal
> more interesting details. The table shows the deltas in percentage to
> the baseline B.
>
> D  R Y
> --
> Timeout  +12%   -16%  -16%
> TailLossProb +28%-7%   -7%
> DSACK_rcvd   +37%-7%   -7%
> Cwnd-undo+16%   -29%  -29%
>
> RTO change affects TLP because TLP will use the min of RTO and TLP
> timer value to arm the probe timer.
>
> The stats indicate that the main culprit of spurious timeouts / rtx is
> the RTO lower-bound. But they also show the RFC RTTVAR averaging is as
> good as current Linux approach.
>
> Given that I would recommend we revise this patch to use the RFC
> averaging but keep existing lower-bound (of RTTVAR to 200ms). We can
> further experiment the lower-bound and change that in a separate
> patch.
Hi I have some update.

I instrumented the kernel to capture the time spent in recovery
(attached). The latency measurement starts when TCP goes into
recovery, triggered by either ACKs or RTOs.  The start time is the
(original) sent time of the first unacked packet. The end time is when
the ACK covers the highest sent sequence when recovery started. The
total latency in usec and count are recorded in MIB_TCPRECOVLAT and
MIB_TCPRECOVCNT. If the connection times out or closes while the
sender was still in recovery, the total latency and count are stored
in MIB_TCPRECOVLAT2 and MIB_TCPRECOVCNT2. This second bucket is to
capture long recovery that led to eventual connection aborts.

Since network stat is usually power distribution, the mean of such
distribution is gonna be dominated by the tail. but the new metrics
still shows very interesting impact of different RTOs. Using the same
table format like my previous email. This table shows the difference
in percentage to the baseline.

  BD  R Y
--
mean TCPRecovLat  3s  -7%   +39% +38%
mean TCPRecovLat252s +1%   -11%  -11%

The new metrics show that lower-bounding RTO to 200ms (D) indeed
lowers the latency. But by my previous analysis, D has a lot more
spurious rtx and TLPs (which the collateral damage on latency is not
captured by these metrics). And note that TLP timer uses the min of
RTO and TLP timeout, so TLP fires 28% more often in (D). Therefore the
latency may be mainly benefited from a faster TLP timer. Nevertheless
the significant impacts on recovery latency do not show up on the
response latency we measured earlier. My conjecture is that only a
small fraction of flows experience losses so even a 40% increase on
average on loss recovery does not move the needle, or the latency

Re: [PATCH net-next v3] tcp: use RFC6298 compliant TCP RTO calculation

[Yuchung Cheng]

On Wed, Jun 22, 2016 at 4:21 AM, Hagen Paul Pfeifer  wrote:
>
> > On June 22, 2016 at 7:53 AM Yuchung Cheng  wrote:
> >
> > Thanks for the patience. I've collected data from some Google Web
> > servers. They serve both a mix of US and SouthAm users using
> > HTTP1 and HTTP2. The traffic is Web browsing (e.g., search, maps,
> > gmails, etc but not Youtube videos). The mean RTT is about 100ms.
> >
> > The user connections were split into 4 groups of different TCP RTO
> > configs. Each group has many millions of connections but the
> > size variation among groups is well under 1%.
> >
> > B: baseline Linux
> > D: this patch
> > R: change RTTYAR averaging as in D, but bound RTO to 1sec per RFC6298
> > Y: change RTTVAR averaging as in D, but bound RTTVAR to 200ms instead (like 
> > B)
> >
> > For mean TCP latency of HTTP responses (first byte sent to last byte
> > acked), B < R < Y < D. But the differences are so insignificant (<1%).
> > The median, 95pctl, and 99pctl has similar indifference. In summary
> > there's hardly visible impact on latency. I also look at only response
> > less than 4KB but do not see a different picture.
> >
> > The main difference is the retransmission rate where R =~ Y < B =~D.
> > R and Y are ~20% lower than B and D. Parsing the SNMP stats reveal
> > more interesting details. The table shows the deltas in percentage to
> > the baseline B.
> >
> > D  R Y
> > --
> > Timeout  +12%   -16%  -16%
> > TailLossProb +28%-7%   -7%
> > DSACK_rcvd   +37%-7%   -7%
> > Cwnd-undo+16%   -29%  -29%
> >
> > RTO change affects TLP because TLP will use the min of RTO and TLP
> > timer value to arm the probe timer.
> >
> > The stats indicate that the main culprit of spurious timeouts / rtx is
> > the RTO lower-bound. But they also show the RFC RTTVAR averaging is as
> > good as current Linux approach.
> >
> > Given that I would recommend we revise this patch to use the RFC
> > averaging but keep existing lower-bound (of RTTVAR to 200ms). We can
> > further experiment the lower-bound and change that in a separate
> > patch.
>
> Great news Yuchung!
>
> Then Daniel will prepare v4 with a min-rto lower bound:
>
> max(RTTVAR, tcp_rto_min_us(struct sock))
>
> Any further suggestions Yuchung, Eric? We will also feed this v4 in our test 
> environment to check the behavior for sender limited, non-continuous flows.
yes a small one: I think the patch should change __tcp_set_rto()
instead of tcp_set_rto() so it applies to recurring timeouts as well.


>
> Hagen

Re: [PATCH net-next v3] tcp: use RFC6298 compliant TCP RTO calculation

[Hagen Paul Pfeifer]

> On June 22, 2016 at 7:53 AM Yuchung Cheng  wrote:
> 
> Thanks for the patience. I've collected data from some Google Web
> servers. They serve both a mix of US and SouthAm users using
> HTTP1 and HTTP2. The traffic is Web browsing (e.g., search, maps,
> gmails, etc but not Youtube videos). The mean RTT is about 100ms.
> 
> The user connections were split into 4 groups of different TCP RTO
> configs. Each group has many millions of connections but the
> size variation among groups is well under 1%.
> 
> B: baseline Linux
> D: this patch
> R: change RTTYAR averaging as in D, but bound RTO to 1sec per RFC6298
> Y: change RTTVAR averaging as in D, but bound RTTVAR to 200ms instead (like B)
> 
> For mean TCP latency of HTTP responses (first byte sent to last byte
> acked), B < R < Y < D. But the differences are so insignificant (<1%).
> The median, 95pctl, and 99pctl has similar indifference. In summary
> there's hardly visible impact on latency. I also look at only response
> less than 4KB but do not see a different picture.
> 
> The main difference is the retransmission rate where R =~ Y < B =~D.
> R and Y are ~20% lower than B and D. Parsing the SNMP stats reveal
> more interesting details. The table shows the deltas in percentage to
> the baseline B.
> 
> D  R Y
> --
> Timeout  +12%   -16%  -16%
> TailLossProb +28%-7%   -7%
> DSACK_rcvd   +37%-7%   -7%
> Cwnd-undo+16%   -29%  -29%
> 
> RTO change affects TLP because TLP will use the min of RTO and TLP
> timer value to arm the probe timer.
> 
> The stats indicate that the main culprit of spurious timeouts / rtx is
> the RTO lower-bound. But they also show the RFC RTTVAR averaging is as
> good as current Linux approach.
> 
> Given that I would recommend we revise this patch to use the RFC
> averaging but keep existing lower-bound (of RTTVAR to 200ms). We can
> further experiment the lower-bound and change that in a separate
> patch.

Great news Yuchung!

Then Daniel will prepare v4 with a min-rto lower bound:

max(RTTVAR, tcp_rto_min_us(struct sock))

Any further suggestions Yuchung, Eric? We will also feed this v4 in our test 
environment to check the behavior for sender limited, non-continuous flows.

Hagen

Re: [PATCH net-next v3] tcp: use RFC6298 compliant TCP RTO calculation

[Yuchung Cheng]

On Fri, Jun 17, 2016 at 11:56 AM, Yuchung Cheng  wrote:
>
> On Fri, Jun 17, 2016 at 11:32 AM, David Miller  wrote:
> >
> > From: Daniel Metz 
> > Date: Wed, 15 Jun 2016 20:00:03 +0200
> >
> > > This patch adjusts Linux RTO calculation to be RFC6298 Standard
> > > compliant. MinRTO is no longer added to the computed RTO, RTO damping
> > > and overestimation are decreased.
> >  ...
> >
> > Yuchung, I assume I am waiting for you to do the testing you said
> > you would do for this patch, right?
> Yes I spent the last two days resolving some unrelated glitches to
> start my testing on Web servers. I should be able to get some results
> over the weekend.
>
> I will test
> 0) current Linux
> 1) this patch
> 2) RFC6298 with min_RTO=1sec
> 3) RFC6298 with minimum RTTVAR of 200ms (so it is more like current
> Linux style of min RTO which only applies to RTTVAR)
>
> and collect the TCP latency (how long to send an HTTP response) and
> (spurious) timeout & retransmission stats.
>
Thanks for the patience. I've collected data from some Google Web
servers. They serve both a mix of US and SouthAm users using
HTTP1 and HTTP2. The traffic is Web browsing (e.g., search, maps,
gmails, etc but not Youtube videos). The mean RTT is about 100ms.

The user connections were split into 4 groups of different TCP RTO
configs. Each group has many millions of connections but the
size variation among groups is well under 1%.

B: baseline Linux
D: this patch
R: change RTTYAR averaging as in D, but bound RTO to 1sec per RFC6298
Y: change RTTVAR averaging as in D, but bound RTTVAR to 200ms instead (like B)

For mean TCP latency of HTTP responses (first byte sent to last byte
acked), B < R < Y < D. But the differences are so insignificant (<1%).
The median, 95pctl, and 99pctl has similar indifference. In summary
there's hardly visible impact on latency. I also look at only response
less than 4KB but do not see a different picture.

The main difference is the retransmission rate where R =~ Y < B =~D.
R and Y are ~20% lower than B and D. Parsing the SNMP stats reveal
more interesting details. The table shows the deltas in percentage to
the baseline B.

D  R Y
--
Timeout  +12%   -16%  -16%
TailLossProb +28%-7%   -7%
DSACK_rcvd   +37%-7%   -7%
Cwnd-undo+16%   -29%  -29%

RTO change affects TLP because TLP will use the min of RTO and TLP
timer value to arm the probe timer.

The stats indicate that the main culprit of spurious timeouts / rtx is
the RTO lower-bound. But they also show the RFC RTTVAR averaging is as
good as current Linux approach.

Given that I would recommend we revise this patch to use the RFC
averaging but keep existing lower-bound (of RTTVAR to 200ms). We can
further experiment the lower-bound and change that in a separate
patch.

Re: [PATCH net-next v3] tcp: use RFC6298 compliant TCP RTO calculation

[Yuchung Cheng]

On Fri, Jun 17, 2016 at 11:32 AM, David Miller  wrote:
>
> From: Daniel Metz 
> Date: Wed, 15 Jun 2016 20:00:03 +0200
>
> > This patch adjusts Linux RTO calculation to be RFC6298 Standard
> > compliant. MinRTO is no longer added to the computed RTO, RTO damping
> > and overestimation are decreased.
>  ...
>
> Yuchung, I assume I am waiting for you to do the testing you said
> you would do for this patch, right?
Yes I spent the last two days resolving some unrelated glitches to
start my testing on Web servers. I should be able to get some results
over the weekend.

I will test
0) current Linux
1) this patch
2) RFC6298 with min_RTO=1sec
3) RFC6298 with minimum RTTVAR of 200ms (so it is more like current
Linux style of min RTO which only applies to RTTVAR)

and collect the TCP latency (how long to send an HTTP response) and
(spurious) timeout & retransmission stats.

I didn't respond to Hagen's email yet b/c I thought data would help
the discussion better :-)

Re: [PATCH net-next v3] tcp: use RFC6298 compliant TCP RTO calculation

[David Miller]

From: Daniel Metz 
Date: Wed, 15 Jun 2016 20:00:03 +0200

> This patch adjusts Linux RTO calculation to be RFC6298 Standard
> compliant. MinRTO is no longer added to the computed RTO, RTO damping
> and overestimation are decreased.
 ...

Yuchung, I assume I am waiting for you to do the testing you said
you would do for this patch, right?

[PATCH net-next v3] tcp: use RFC6298 compliant TCP RTO calculation

[Daniel Metz]

From: Daniel Metz 

This patch adjusts Linux RTO calculation to be RFC6298 Standard
compliant. MinRTO is no longer added to the computed RTO, RTO damping
and overestimation are decreased.

In RFC 6298 Standard TCP Retransmission Timeout (RTO) calculation the
calculated RTO is rounded up to the Minimum RTO (MinRTO), if it is
less.  The Linux implementation as a discrepancy to the Standard
basically adds the defined MinRTO to the calculated RTO. When
comparing both approaches, the Linux calculation seems to perform
worse for sender limited TCP flows like Telnet, SSH or constant bit
rate encoded transmissions, especially for Round Trip Times (RTT) of
50ms to 800ms.

Compared to the Linux implementation the RFC 6298 proposed RTO
calculation performs better and more precise in adapting to current
network characteristics. Extensive measurements for bulk data did not
show a negative impact of the adjusted calculation.

Exemplary performance comparison for sender-limited-flows:

- Rate: 10Mbit/s
- Delay: 200ms, Delay Variation: 10ms
- Time between each scheduled segment: 1s
- Amount Data Segments: 300
- Mean of 11 runs

 Mean Response Waiting Time [milliseconds]

PER [%] |   0.5  11.5  2  3  5  7 10
+---
old | 206.4  208.6  218.0  218.6  227.2  249.3  274.7  308.2
new | 203.9  206.0  207.0  209.9  217.3  225.6  238.7  259.1

Detailed analysis:
https://docs.google.com/document/d/1pKmPfnQb6fDK4qpiNVkN8cQyGE4wYDZukcuZfR-BnnM/

Reasoning for historic design:
Sarolahti, P.; Kuznetsov, A. (2002). Congestion Control in Linux TCP.
Conference Paper. Proceedings of the FREENIX Track. 2002 USENIX Annual
https://www.cs.helsinki.fi/research/iwtcp/papers/linuxtcp.pdf

Signed-off-by: Hagen Paul Pfeifer 
Signed-off-by: Daniel Metz 
Cc: Eric Dumazet 
Cc: Yuchung Cheng 
---

 v3:
  - remove mdev_us

 v2:
  - Using the RFC 6298 compliant implementation, the tcp_sock struct variable
u32 mdev_max_us becomes obsolete and consequently is being removed.
  - Add reference to Kuznetsov paper


 include/linux/tcp.h  |  2 --
 net/ipv4/tcp.c   |  3 +-
 net/ipv4/tcp_input.c | 72 ++--
 net/ipv4/tcp_metrics.c   |  5 ++--
 net/ipv4/tcp_minisocks.c |  1 -
 5 files changed, 18 insertions(+), 65 deletions(-)

diff --git a/include/linux/tcp.h b/include/linux/tcp.h
index 7be9b12..3128eb1 100644
--- a/include/linux/tcp.h
+++ b/include/linux/tcp.h
@@ -230,8 +230,6 @@ struct tcp_sock {
 
 /* RTT measurement */
u32 srtt_us;/* smoothed round trip time << 3 in usecs */
-   u32 mdev_us;/* medium deviation */
-   u32 mdev_max_us;/* maximal mdev for the last rtt period */
u32 rttvar_us;  /* smoothed mdev_max*/
u32 rtt_seq;/* sequence number to update rttvar */
struct rtt_meas {
diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 5c7ed14..4a7597c 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -386,7 +386,6 @@ void tcp_init_sock(struct sock *sk)
INIT_LIST_HEAD(>tsq_node);
 
icsk->icsk_rto = TCP_TIMEOUT_INIT;
-   tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
tp->rtt_min[0].rtt = ~0U;
 
/* So many TCP implementations out there (incorrectly) count the
@@ -2703,7 +2702,7 @@ void tcp_get_info(struct sock *sk, struct tcp_info *info)
info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
info->tcpi_rtt = tp->srtt_us >> 3;
-   info->tcpi_rttvar = tp->mdev_us >> 2;
+   info->tcpi_rttvar = tp->rttvar_us >> 2;
info->tcpi_snd_ssthresh = tp->snd_ssthresh;
info->tcpi_snd_cwnd = tp->snd_cwnd;
info->tcpi_advmss = tp->advmss;
diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
index 94d4aff..279f5f7 100644
--- a/net/ipv4/tcp_input.c
+++ b/net/ipv4/tcp_input.c
@@ -680,8 +680,7 @@ static void tcp_event_data_recv(struct sock *sk, struct 
sk_buff *skb)
 /* Called to compute a smoothed rtt estimate. The data fed to this
  * routine either comes from timestamps, or from segments that were
  * known _not_ to have been retransmitted [see Karn/Partridge
- * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
- * piece by Van Jacobson.
+ * Proceedings SIGCOMM 87].
  * NOTE: the next three routines used to be one big routine.
  * To save cycles in the RFC 1323 implementation it was better to break
  * it up into three procedures. -- erics
@@ -692,59 +691,19 @@ static void tcp_rtt_estimator(struct sock *sk, long 
mrtt_us)
long m = mrtt_us; /* RTT */
u32 srtt = tp->srtt_us;
 
-   /*  The following amusing code comes from Jacobson's
-*  article in SIGCOMM '88.  Note that rtt and mdev
-*  are scaled