Re: [RFC] tcp: Delay sending non-probes for RFC4821 mtu probing

On Mon, Apr 26, 2021 at 1:09 PM Leonard Crestez [off-list ref] wrote:

On 26.04.2021 18:59, Neal Cardwell wrote:

quoted

On Sun, Apr 25, 2021 at 10:34 PM Leonard Crestez [off-list ref] wrote:

quoted

On 4/21/21 3:47 PM, Neal Cardwell wrote:

quoted

On Wed, Apr 21, 2021 at 6:21 AM Leonard Crestez [off-list ref] wrote:

quoted

quoted

quoted

If the goal is to increase the frequency of PMTU probes, which seems
like a valid goal, I would suggest that we rethink the Linux heuristic
for triggering PMTU probes in the light of the fact that the loss
detection mechanism is now RACK-TLP, which provides quick recovery in
a much wider variety of scenarios.

quoted

You mention:

quoted

Linux waits for probe_size + (1 + retries) * mss_cache to be available

The code in question seems to be:

    size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
How about just changing this to:

    size_needed = probe_size + tp->mss_cache;

The rationale would be that if that amount of data is available, then
the sender can send one probe and one following current-mss-size
packet. If the path MTU has not increased to allow the probe of size
probe_size to pass through the network, then the following
current-mss-size packet will likely pass through the network, generate
a SACK, and trigger a RACK fast recovery 1/4*min_rtt later, when the
RACK reorder timer fires.

This appears to almost work except it stalls after a while. I spend some
time investigating it and it seems that cwnd is shrunk on mss increases
and does not go back up. This causes probes to be skipped because of a
"snd_cwnd < 11" condition.

I don't undestand where that magical "11" comes from, could that be
shrunk. Maybe it's meant to only send probes when the cwnd is above the
default of 10? Then maybe mtu_probe_success shouldn't shrink mss below
what is required for an additional probe, or at least round-up.

The shrinkage of cwnd is a problem with this "short probes" approach
because tcp_is_cwnd_limited returns false because tp->max_packets_out is
smaller (4). With longer probes tp->max_packets_out is larger (6) so
tcp_is_cwnd_limited returns true even for a cwnd of 10.

I'm testing using namespace-to-namespace loopback so my delays are close
to zero. I tried to introduce an artificial delay of 30ms (using tc
netem) and it works but 20ms does not.

I agree the magic 11 seems outdated and unnecessarily high, given RACK-TLP.

I think it would be fine to change the magic 11 to a magic
(TCP_FASTRETRANS_THRESH+1), aka 3+1=4:

   - tp->snd_cwnd < 11 ||
   + p->snd_cwnd < (TCP_FASTRETRANS_THRESH + 1) ||

As long as the cwnd is >= TCP_FASTRETRANS_THRESH+1 then the sender
should usually be able to send the 1 probe packet and then 3
additional packets beyond the probe, and in the common case (with no
reordering) then with failed probes this should allow the sender to
quickly receive 3 SACKed segments and enter fast recovery quickly.
Even if the sender doesn't have 3 additional packets, or if reordering
has been detected, then RACK-TLP should be able to start recovery
quickly (5/4*RTT if there is at least one SACK, or 2*RTT for a TLP if
there is no SACK).

As far as I understand tp->reordering is a dynamic evaluation of the
fastretrans threshold to deal with environments with lots of reordering.
Your suggestion seems equivalent to the current size_needed calculation
except using packets instead of bytes.

Wouldn't it be easier to drop the "11" check and just verify that
size_needed fits into cwnd as bytes?

Yes, that sounds good to me (dropping the "11" check in favor of
verifying that size_needed fits into the cwnd).

neal