Re: [PATCH net-next] fq_codel: avoid under-utilization with ce_threshold at low link rates

On Fri, Oct 29, 2021 at 6:54 AM Neal Cardwell [off-list ref] wrote:

On Thu, Oct 28, 2021 at 3:15 PM Asad Sajjad Ahmed [off-list ref] wrote:

quoted

Commit "fq_codel: generalise ce_threshold marking for subset of traffic"
[1] enables ce_threshold to be used in the Internet, not just in data
centres.

Because ce_threshold is in time units, it can cause poor utilization at
low link rates when it represents <1 packet.
E.g., if link rate <12Mb/s ce_threshold=1ms is <1500B packet.

So, suppress ECN marking unless the backlog is also > 1 MTU.

A similar patch to [1] was tested on an earlier kernel, and a similar
one-packet check prevented poor utilization at low link rates [2].

[1] commit dfcb63ce1de6 ("fq_codel: generalise ce_threshold marking for subset of traffic")

[2] See right hand column of plots at the end of:
https://bobbriscoe.net/projects/latency/dctth_journal_draft20190726.pdf

Signed-off-by: Asad Sajjad Ahmed <redacted>
Signed-off-by: Olga Albisser <redacted>
---
 include/net/codel_impl.h | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/include/net/codel_impl.h b/include/net/codel_impl.h
index 137d40d8cbeb..4e3e8473e776 100644
--- a/include/net/codel_impl.h
+++ b/include/net/codel_impl.h

@@ -248,7 +248,8 @@ static struct sk_buff *codel_dequeue(void *ctx,
                                                    vars->rec_inv_sqrt);
        }
 end:
-       if (skb && codel_time_after(vars->ldelay, params->ce_threshold)) {
+       if (skb && codel_time_after(vars->ldelay, params->ce_threshold) &&
+           *backlog > params->mtu) {

I think this idea would apply to codel quite well.  (This helper is
common to codel and fq_codel)

But with fq_codel my thoughts are:

*backlog is the backlog of the qdisc, not the backlog for the flow,
and it includes the packet currently being removed from the queue.

Setting ce_threshold to 1ms while the link rate is 12Mbs sounds
misconfiguration to me.

Even if this flow has to transmit one tiny packet every minute, it
will get CE mark
just because at least one packet from an elephant flow is currently
being sent to the wire.

BQL won't prevent that at least one packet is being processed while
the tiny packet
is coming into fq_codel qdisc.

vars->ldelay = now - skb_time_func(skb);

For tight ce_threshold, vars->ldelay would need to be replaced by

now - (time of first codel_dequeue() after this skb has been queued).
This seems a bit hard to implement cheaply.

quoted

                bool set_ce = true;

                if (params->ce_threshold_mask) {
--

Sounds like a good idea, and looks good to me.

Acked-by: Neal Cardwell <ncardwell@google.com>

Eric, what do you think?

neal