On Tue, Jun 21, 2016 at 10:53 PM, Yuchung Cheng [off-list ref] wrote:

On Fri, Jun 17, 2016 at 11:56 AM, Yuchung Cheng [off-list ref] wrote:

quoted

On Fri, Jun 17, 2016 at 11:32 AM, David Miller [off-list ref] wrote:

quoted

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

quoted

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.

                                      B        D      R     Y
------------------------------------------------------------------------------
mean TCPRecovLat      3s      -7%   +39% +38%
mean TCPRecovLat2    52s     +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
reduction was cancelled by the latency increase of spurious timeouts
and CC reactions.

hmm it almost seen that the current code is the right balance. We may
need more investigation.

Daniel and Hagen, could you try my instrumentation patch on your testbed?