[Cake] a little bit of cake testing

[Cake] a little bit of cake testing

[Dave Taht]

[-- Attachment #1: Type: text/plain, Size: 1049 bytes --]

My network here is in flux (new modem, signal strength problems, cable
problems) and toke's testbed is presently doing wifi work, so I did a
quick mod to fishcake to make it do linux 4.0 (note I am not sure if
this was a 4.1 or a 4.0 change) - attached. (and we lose a few cake
options due to me not grokking the new API)

Cake did well on this, but the behavior at the tail end of the test
was disturbing:

http://www.dslreports.com/speedtest/611312

Need to do some work to emulate the dslreports tests.

And it dropped  LOT more packets than fq_codel did. fq_codel marked 33
packets for it's result, cake marked 600 and dropped 200, for its.

I did some rrul testing as well, but was fighting with a modem that
used to get 140Mbits, and now only gets 70mbits, and behaves very
differently overall with pfifo_fast than i had ever seen before. And
along the way snapon got upgraded a bit too... sigh...

-- 
Dave Täht
What will it take to vastly improve wifi for everyone?
https://plus.google.com/u/0/explore/makewififast

[-- Attachment #2: 0001-Make-compile-on-linux-4.0-and-later.patch --]
[-- Type: text/x-patch, Size: 2099 bytes --]

From ac49485b29519fb1a4a88bc07e715ea232058397 Mon Sep 17 00:00:00 2001
From: Dave Taht <dave.taht@bufferbloat.net>
Date: Thu, 4 Jun 2015 10:04:46 -0700
Subject: [PATCH] Make compile on linux 4.0 and later

---
 sch_cake.c | 12 ++++++++++--
 1 file changed, 10 insertions(+), 2 deletions(-)

diff --git a/sch_cake.c b/sch_cake.c
index 242333c..538d090 100644
--- a/sch_cake.c
+++ b/sch_cake.c
@@ -49,10 +49,12 @@
 #include <linux/vmalloc.h>
 #include <linux/reciprocal_div.h>
 #include <net/netlink.h>
+#include <linux/version.h>
 #include "pkt_sched.h"
+#if LINUX_VERSION_CODE < KERNEL_VERSION(4,0,0)
 #include <net/flow_keys.h>
+#endif
 #include "codel5.h"
-
 /* The CAKE Principle:
  *                 (or, how to have your cake and eat it too)
  *
@@ -205,12 +207,15 @@ enum {
 static inline unsigned int
 cake_fqcd_hash(struct cake_fqcd_sched_data *q, const struct sk_buff *skb, int flow_mode)
 {
+#if LINUX_VERSION_CODE < KERNEL_VERSION(4,0,0)
     struct flow_keys keys;
+#endif
     u32 hash, reduced_hash;
 
 	if(unlikely(flow_mode == CAKE_FLOW_NONE || q->flows_cnt < CAKE_SET_WAYS))
 		return 0;
 
+#if LINUX_VERSION_CODE < KERNEL_VERSION(4,0,0)
 	skb_flow_dissect(skb, &keys);
 
 	if(flow_mode != CAKE_FLOW_ALL) {
@@ -227,6 +232,9 @@ cake_fqcd_hash(struct cake_fqcd_sched_data *q, const struct sk_buff *skb, int fl
 	hash = jhash_3words((__force u32)keys.dst,
 						(__force u32)keys.src ^ keys.ip_proto,
 						(__force u32)keys.ports, q->perturbation);
+#else
+	hash = skb_get_hash_perturb(skb, q->perturbation);
+#endif
 
 	reduced_hash = reciprocal_scale(hash, q->flows_cnt);
 
@@ -416,7 +424,7 @@ static int cake_enqueue(struct sk_buff *skb, struct Qdisc *sch)
 	 * Split GSO aggregates if they're likely to impair flow isolation
 	 * or if we need to know individual packet sizes for framing overhead.
 	 */
-	if(unlikely((len * max(fqcd->flow_count, 1)) > q->peel_threshold && skb_is_gso(skb)))
+	if(((len * max((fqcd->flow_count), 1)) > q->peel_threshold && skb_is_gso(skb)))
 	{
 		struct sk_buff *segs, *nskb;
 		netdev_features_t features = netif_skb_features(skb);
-- 
2.1.0

Re: [Cake] a little bit of cake testing

[Dave Taht]

I did a bit of setup on my connection to give me a decent rrul result.
(I was not rate limiting inbound enough) The duration of the initial
load spike is much less pronounced than the fq_codel result. I note
that I have offloads still on, so I imagine fq_codel is getting
tweaked by that....

but ingress gets way out of hand later on in this test. I can make an
argument for decay (count/2) being far too aggressive. In fact, even
count - 2 seemed too much in older testing I had done with other
variants. I would certainly like to get a feel for when and where the
three parts of codel are kicking in in various workloads.

http://www.dslreports.com/speedtest/613646

I really need to get to where I can quickly get to a blog entry on
this stuff, and back to comprehensive, controlled testing.





On Thu, Jun 4, 2015 at 10:11 AM, Dave Taht <dave.taht@gmail.com> wrote:
> My network here is in flux (new modem, signal strength problems, cable
> problems) and toke's testbed is presently doing wifi work, so I did a
> quick mod to fishcake to make it do linux 4.0 (note I am not sure if
> this was a 4.1 or a 4.0 change) - attached. (and we lose a few cake
> options due to me not grokking the new API)
>
> Cake did well on this, but the behavior at the tail end of the test
> was disturbing:
>
> http://www.dslreports.com/speedtest/611312
>
> Need to do some work to emulate the dslreports tests.
>
> And it dropped  LOT more packets than fq_codel did. fq_codel marked 33
> packets for it's result, cake marked 600 and dropped 200, for its.
>
> I did some rrul testing as well, but was fighting with a modem that
> used to get 140Mbits, and now only gets 70mbits, and behaves very
> differently overall with pfifo_fast than i had ever seen before. And
> along the way snapon got upgraded a bit too... sigh...
>
> --
> Dave Täht
> What will it take to vastly improve wifi for everyone?
> https://plus.google.com/u/0/explore/makewififast



-- 
Dave Täht
What will it take to vastly improve wifi for everyone?
https://plus.google.com/u/0/explore/makewififast

Re: [Cake] a little bit of cake testing

[Dave Taht]

not very well controlled dataset at:
http://snapon.lab.bufferbloat.net/~d/fishcake/

On Thu, Jun 4, 2015 at 11:10 AM, Dave Taht <dave.taht@gmail.com> wrote:
> I did a bit of setup on my connection to give me a decent rrul result.
> (I was not rate limiting inbound enough) The duration of the initial
> load spike is much less pronounced than the fq_codel result. I note
> that I have offloads still on, so I imagine fq_codel is getting
> tweaked by that....
>
> but ingress gets way out of hand later on in this test. I can make an
> argument for decay (count/2) being far too aggressive. In fact, even
> count - 2 seemed too much in older testing I had done with other
> variants. I would certainly like to get a feel for when and where the
> three parts of codel are kicking in in various workloads.
>
> http://www.dslreports.com/speedtest/613646
>
> I really need to get to where I can quickly get to a blog entry on
> this stuff, and back to comprehensive, controlled testing.
>
>
>
>
>
> On Thu, Jun 4, 2015 at 10:11 AM, Dave Taht <dave.taht@gmail.com> wrote:
>> My network here is in flux (new modem, signal strength problems, cable
>> problems) and toke's testbed is presently doing wifi work, so I did a
>> quick mod to fishcake to make it do linux 4.0 (note I am not sure if
>> this was a 4.1 or a 4.0 change) - attached. (and we lose a few cake
>> options due to me not grokking the new API)
>>
>> Cake did well on this, but the behavior at the tail end of the test
>> was disturbing:
>>
>> http://www.dslreports.com/speedtest/611312
>>
>> Need to do some work to emulate the dslreports tests.
>>
>> And it dropped  LOT more packets than fq_codel did. fq_codel marked 33
>> packets for it's result, cake marked 600 and dropped 200, for its.
>>
>> I did some rrul testing as well, but was fighting with a modem that
>> used to get 140Mbits, and now only gets 70mbits, and behaves very
>> differently overall with pfifo_fast than i had ever seen before. And
>> along the way snapon got upgraded a bit too... sigh...
>>
>> --
>> Dave Täht
>> What will it take to vastly improve wifi for everyone?
>> https://plus.google.com/u/0/explore/makewififast
>
>
>
> --
> Dave Täht
> What will it take to vastly improve wifi for everyone?
> https://plus.google.com/u/0/explore/makewififast



-- 
Dave Täht
What will it take to vastly improve wifi for everyone?
https://plus.google.com/u/0/explore/makewififast

Re: [Cake] a little bit of cake testing

[Jonathan Morton]

[-- Attachment #1: Type: text/plain, Size: 3370 bytes --]

In summary, the new Codel behaviour works better on egress but worse on
ingress. That's reasonable - on ingress the queue always appears to build
slowly, so the new scheme always triggers late, and is more likely to
return aggressively to a lower signalling rate. I think we're going to need
a special mode for ingress.

However, I think a good deal of your poor results are due to not properly
matching the actual link bandwidth any more. I would suspect trouble in the
cable rather than the modem, given the symptoms.

- Jonathan Morton
 On 4 Jun 2015 21:11, "Dave Taht" <dave.taht@gmail.com> wrote:

> not very well controlled dataset at:
> http://snapon.lab.bufferbloat.net/~d/fishcake/
>
> On Thu, Jun 4, 2015 at 11:10 AM, Dave Taht <dave.taht@gmail.com> wrote:
> > I did a bit of setup on my connection to give me a decent rrul result.
> > (I was not rate limiting inbound enough) The duration of the initial
> > load spike is much less pronounced than the fq_codel result. I note
> > that I have offloads still on, so I imagine fq_codel is getting
> > tweaked by that....
> >
> > but ingress gets way out of hand later on in this test. I can make an
> > argument for decay (count/2) being far too aggressive. In fact, even
> > count - 2 seemed too much in older testing I had done with other
> > variants. I would certainly like to get a feel for when and where the
> > three parts of codel are kicking in in various workloads.
> >
> > http://www.dslreports.com/speedtest/613646
> >
> > I really need to get to where I can quickly get to a blog entry on
> > this stuff, and back to comprehensive, controlled testing.
> >
> >
> >
> >
> >
> > On Thu, Jun 4, 2015 at 10:11 AM, Dave Taht <dave.taht@gmail.com> wrote:
> >> My network here is in flux (new modem, signal strength problems, cable
> >> problems) and toke's testbed is presently doing wifi work, so I did a
> >> quick mod to fishcake to make it do linux 4.0 (note I am not sure if
> >> this was a 4.1 or a 4.0 change) - attached. (and we lose a few cake
> >> options due to me not grokking the new API)
> >>
> >> Cake did well on this, but the behavior at the tail end of the test
> >> was disturbing:
> >>
> >> http://www.dslreports.com/speedtest/611312
> >>
> >> Need to do some work to emulate the dslreports tests.
> >>
> >> And it dropped  LOT more packets than fq_codel did. fq_codel marked 33
> >> packets for it's result, cake marked 600 and dropped 200, for its.
> >>
> >> I did some rrul testing as well, but was fighting with a modem that
> >> used to get 140Mbits, and now only gets 70mbits, and behaves very
> >> differently overall with pfifo_fast than i had ever seen before. And
> >> along the way snapon got upgraded a bit too... sigh...
> >>
> >> --
> >> Dave Täht
> >> What will it take to vastly improve wifi for everyone?
> >> https://plus.google.com/u/0/explore/makewififast
> >
> >
> >
> > --
> > Dave Täht
> > What will it take to vastly improve wifi for everyone?
> > https://plus.google.com/u/0/explore/makewififast
>
>
>
> --
> Dave Täht
> What will it take to vastly improve wifi for everyone?
> https://plus.google.com/u/0/explore/makewififast
> _______________________________________________
> Cake mailing list
> Cake@lists.bufferbloat.net
> https://lists.bufferbloat.net/listinfo/cake
>

[-- Attachment #2: Type: text/html, Size: 4710 bytes --]