Hi All,
> On Nov 28, 2017, at 23:37, Dave Taht wrote:
>
>
> A flag day here is feasible. I will fiddle along the lines you describe.
>
> As for other flag days...
>
> I'm toying with the idea of fixing xstats in a separate branch. I really
> hate the idea of breaking backward
Dave Taht writes:
> A flag day here is feasible. I will fiddle along the lines you
> describe.
FWIW I don't think the history is that bad. Sure, there are a bunch of
merge commits, but picking out the real ones is not that difficult
(unless you are using the github web interface,
> On Nov 29, 2017, at 9:19 AM, Pete Heist wrote:
>
>> On Nov 29, 2017, at 4:42 AM, Georgios Amanakis wrote:
>>
>> @Pete I think you need to start netserver on the client first (in your case,
>> you are running flent on the server): "ip netns exec
With ack-filter-aggressive I get results similar to rrul1 in ~10% of the
runs, 90% are similar to rrul2.
With ack-filter I haven't managed yet to get a rrul1 result, all of them
are similar to rrul2.
Will experiment more today.
George
On Tue, Nov 28, 2017 at 11:01 PM, Dave Taht
Georgios Amanakis writes:
> I am troubled by the number of data points flent reports for some pings and
> uploads in this setup. A typical ack-filter result, similar to rrul2 I
> posted before, looks like this:
> Summary of rrul test run 'rrul_cakeeth_ds3_900mbit_45mbit_ack'
> On Nov 29, 2017, at 3:50 PM, Toke Høiland-Jørgensen wrote:
>
> Georgios Amanakis > writes:
>
>> I am troubled by the number of data points flent reports for some pings and
>> uploads in this setup. A typical ack-filter result,
> (That was also informative for me about how netperf decides when to
> emit a data point…)
In that case I can add that the stated reason for this way of doing
things is performance (i.e., emitting data points should not interfere
with transfer performance). This is mostly an issue on systems
> On Nov 29, 2017, at 4:44 PM, Toke Høiland-Jørgensen wrote:
>
>> (That was also informative for me about how netperf decides when to
>> emit a data point…)
>
> In that case I can add that the stated reason for this way of doing
> things is performance (i.e., emitting data points
I am troubled by the number of data points flent reports for some pings and
uploads in this setup. A typical ack-filter result, similar to rrul2 I
posted before, looks like this:
Summary of rrul test run 'rrul_cakeeth_ds3_900mbit_45mbit_ack' (at
2017-11-29 14:37:5
5.719180):
I did some more testing. Same setup as before, I varied the amount of delay:
server--delay--mbox--client
netserver Xms/Xms 45/900mbit
Cake config:
qdisc cake 801b: dev mbox.l root refcnt 2 bandwidth 45Mbit diffserv3
Sebastian Moeller writes:
> Hi All,
>
>
>> On Nov 28, 2017, at 23:37, Dave Taht wrote:
>>
>>
>> A flag day here is feasible. I will fiddle along the lines you describe.
>>
>> As for other flag days...
>>
>> I'm toying with the idea of fixing xstats in a
My fault, I had forgotten to include ~/go/bin in PATH.
New files:
delay 100ms (rtt) flent: https://drive.google.com/open?
id=13HAAvDtkBCPNu5FC61Cge_1vcd5LXuWY
delay 100ms (rtt) stat: https://drive.google.com/open?id=1THM-
F0MiuCntMz7lx3m7vA1FvHYZSNHC
On Wed, Nov 29, 2017 at 11:40 AM, Toke
On Wed, Nov 29, 2017 at 5:24 PM Georgios Amanakis
wrote:
> I just installed flent-git (thanks for the aur) and irtt, was more
> straightforward than I expected :)
> In my setup netserver runs on server, flent runs on client. Where should I
> run irtt server?
>
irtt server
-- Forwarded message --
From: Georgios Amanakis
Date: Wed, Nov 29, 2017 at 12:50 PM
Subject: Re: [Cake] cake flenter results round 2
To: Dave Taht
To avoid a misunderstanding, the delay parameter in both:
"ip netns exec delay tc qdisc replace
Also, it is easier (for me at least) to download a tarball of all your
results for a given run, rather than each one individually.
I am thinking we can rename ack-filter-aggressive to
ack-filter-too-damn-aggressive.
Dave Taht writes:
> I just want to verify that you increased
Georgios Amanakis writes:
> Results with irtt/flent-git. Same setup as before:
Looks like that is still using netperf for the UDP measurements. Flent
does some sanity checks on irtt before using it, which may be failing.
Running flent with -v should give some hints...
@Pete @Toke just saw your responses. Thank you very much for the
explanation. I will give irtt and flent-git a try, and maybe build an aur
package for archlinux.
George
On Wed, Nov 29, 2017 at 11:08 AM, Pete Heist wrote:
>
> > On Nov 29, 2017, at 4:44 PM, Toke
Georgios Amanakis writes:
> @Pete @Toke just saw your responses. Thank you very much for the
> explanation. I will give irtt and flent-git a try, and maybe build an
> aur package for archlinux.
There's already an AUR package for flent-git; haven't gotten around to
making
Very interesting. The aggressive-ack-filter seems to mess throughput up during
"boost phase" compared to ack-filter less agreesive.
I should say that my most common heavy upload/download usage has an RTT of 88
msec. (Boston - AWS west).
Personal heavy upload/download usage is has RTT of
dpr...@reed.com writes:
> BTW, one annoying thing about flent is that netperf is not available
> in fedora, rhel, centos as a package. So I have to install it and it
> doesn't just plug in smoothly. That's just my convenience issue. I
> don't use netperf myself on any of my systems. iperf and
On Wed, Nov 29, 2017 at 10:21 AM, Juliusz Chroboczek wrote:
>> The better solution would of course be to have the TCP peeps change the
>> way TCP works so that it sends fewer ACKs.
>
> Which tends to perturb the way the TCP self-clocking feedback loop works,
> and to break Nagle.
there is no performance impact of using really high values for netem
limit. Stick with 10. :)
(well, there is a cache impact, but that's the cost of correct simulation)
On Wed, Nov 29, 2017 at 10:19 AM, Georgios Amanakis wrote:
> I completely neglected this. It turns
On Wed, Nov 29, 2017 at 10:28 AM, Juliusz Chroboczek wrote:
>> Recently Ryan Mounce added ack filtering cabilities to the cake qdisc.
>> The benefits were pretty impressive at a 50x1 Down/Up ratio:
>
> If I read this posting right, you're only measuring bulk performance.
> What
23 matches
Mail list logo