[LibreQoS] [Starlink] [Rpm] Researchers Seeking Probe Volunteers in USA
dickroy at alum.mit.edu
Thu Jan 12 16:57:32 EST 2023
Nothing earth-shaking :-)
From: Starlink [mailto:starlink-bounces at lists.bufferbloat.net] On Behalf Of
Robert McMahon via Starlink
Sent: Thursday, January 12, 2023 9:50 AM
To: Sebastian Moeller
Cc: Dave Taht via Starlink; mike.reynolds at netforecast.com; libreqos; David
P. Reed; Rpm; bloat
Subject: Re: [Starlink] [Rpm] Researchers Seeking Probe Volunteers in USA
You make a good point. What I did was issue a warning if the tool found it
was being CPU limited vs i/o limited. This indicates the i/o test likely is
inaccurate from an i/o perspective, and the results are suspect. It does
this crudely by comparing the cpu thread doing stats against the traffic
threads doing i/o, which thread is waiting on the others. There is no
attempt to assess the cpu load itself. So it's designed with a singular
purpose of making sure i/o threads only block on syscalls of write and read.
I probably should revisit this both in design and implementation. Thanks for
bringing it up and all input is truly appreciated.
On Jan 12, 2023, at 12:14 AM, Sebastian Moeller <moeller0 at gmx.de> wrote:
On Jan 11, 2023, at 21:09, rjmcmahon <rjmcmahon at rjmcmahon.com> wrote:
Iperf 2 is designed to measure network i/o. Note: It doesn't have to move
large amounts of data. It can support data profiles that don't drive TCP's
CCA as an example.
Two things I've been asked for and avoided:
1) Integrate clock sync into iperf's test traffic
[SM] This I understand, measurement conditions can be unsuited for tight
2) Measure and output CPU usages
[SM] This one puzzles me, as far as I understand the only way to properly
diagnose network issues is to rule out other things like CPU overload that
can have symptoms similar to network issues. As an example, the cake qdisc
will if CPU cycles become tight first increases its internal queueing and
jitter (not consciously, it is just an observation that once cake does not
get access to the CPU as timely as it wants, queuing latency and variability
increases) and then later also shows reduced throughput, so similar things
that can happen along an e2e network path for completely different reasons,
e.g. lower level retransmissions or a variable rate link. So i would think
that checking the CPU load at least coarse would be within the scope of
network testing tools, no?
I think both of these are outside the scope of a tool designed to test
network i/o over sockets, rather these should be developed & validated
independently of a network i/o tool.
Clock error really isn't about amount/frequency of traffic but rather
getting a periodic high-quality reference. I tend to use GPS pulse per
second to lock the local system oscillator to. As David says, most every
modern handheld computer has the GPS chips to do this already. So to me it
seems more of a policy choice between data center operators and device mfgs
and less of a technical issue.
Yall can call me crazy if you want.. but... see below [RWG]
On Jan 9, 2023, at 20:13, rjmcmahon via Starlink
<starlink at lists.bufferbloat.net> wrote:
My biggest barrier is the lack of clock sync by the devices, i.e. very
limited support for PTP in data centers and in end devices. This limits the
ability to measure one way delays (OWD) and most assume that OWD is 1/2 and
RTT which typically is a mistake. We know this intuitively with airplane
flight times or even car commute times where the one way time is not 1/2 a
round trip time. Google maps & directions provide a time estimate for the
one way link. It doesn't compute a round trip and divide by two.
For those that can get clock sync working, the iperf 2 --trip-times options
[SM] +1; and yet even with unsynchronized clocks one can try to measure
how latency changes under load and that can be done per direction. Sure this
is far inferior to real reliably measured OWDs, but if life/the internet
deals you lemons....
[RWG] iperf2/iperf3, etc are already moving large amounts of data
back and forth, for that matter any rate test, why not abuse some of
that data and add the fundemental NTP clock sync data and
bidirectionally pass each others concept of "current time". IIRC (its
been 25 years since I worked on NTP at this level) you *should* be
able to get a fairly accurate clock delta between each end, and then
use that info and time stamps in the data stream to compute OWD's.
You need to put 4 time stamps in the packet, and with that you can
enable the measurement of end to end write to read latencies (client and
server clocks must be synchronized)
enable the measurement of the wall clock difference between sender and
I have many kvetches about the new latency under load tests being
designed and distributed over the past year. I am delighted! that they
are happening, but most really need third party evaluation, and
calibration, and a solid explanation of what network pathologies they
do and don't cover. Also a RED team attitude towards them, as well as
thinking hard about what you are not measuring (operations research).
I actually rather love the new cloudflare speedtest, because it tests
a single TCP connection, rather than dozens, and at the same time folk
are complaining that it doesn't find the actual "speed!". yet... the
test itself more closely emulates a user experience than speedtest.net
does. I am personally pretty convinced that the fewer numbers of flows
that a web page opens improves the likelihood of a good user
experience, but lack data on it.
To try to tackle the evaluation and calibration part, I've reached out
to all the new test designers in the hope that we could get together
and produce a report of what each new test is actually doing. I've
tweeted, linked in, emailed, and spammed every measurement list I know
of, and only to some response, please reach out to other test designer
folks and have them join the rpm email list?
My principal kvetches in the new tests so far are:
0) None of the tests last long enough.
Ideally there should be a mode where they at least run to "time of
first loss", or periodically, just run longer than the
industry-stupid^H^H^H^H^H^Hstandard 20 seconds. There be dragons
there! It's really bad science to optimize the internet for 20
seconds. It's like optimizing a car, to handle well, for just 20
1) Not testing up + down + ping at the same time
None of the new tests actually test the same thing that the infamous
rrul test does - all the others still test up, then down, and ping. It
was/remains my hope that the simpler parts of the flent test suite -
such as the tcp_up_squarewave tests, the rrul test, and the rtt_fair
tests would provide calibration to the test designers.
we've got zillions of flent results in the archive published here:
ps. Misinformation about iperf 2 impacts my ability to do this.
The new tests have all added up + ping and down + ping, but not up +
down + ping. Why??
The behaviors of what happens in that case are really non-intuitive, I
know, but... it's just one more phase to add to any one of those new
tests. I'd be deliriously happy if someone(s) new to the field
started doing that, even optionally, and boggled at how it defeated
Among other things that would show...
It's the home router industry's dirty secret than darn few "gigabit"
home routers can actually forward in both directions at a gigabit. I'd
like to smash that perception thoroughly, but given our starting point
is a gigabit router was a "gigabit switch" - and historically been
something that couldn't even forward at 200Mbit - we have a long way
to go there.
Only in the past year have non-x86 home routers appeared that could
actually do a gbit in both directions.
2) Few are actually testing within-stream latency
Apple's rpm project is making a stab in that direction. It looks
highly likely, that with a little more work, crusader and
go-responsiveness can finally start sampling the tcp RTT, loss and
markings, more directly. As for the rest... sampling TCP_INFO on
windows, and Linux, at least, always appeared simple to me, but I'm
discovering how hard it is by delving deep into the rust behind
the goresponsiveness thing is also IMHO running WAY too many streams
at the same time, I guess motivated by an attempt to have the test
B) To try and tackle the validation problem:ps. Misinformation about iperf
2 impacts my ability to do this.
In the libreqos.io project we've established a testbed where tests can
be plunked through various ISP plan network emulations. It's here:
https://payne.taht.net (run bandwidth test for what's currently hooked
We could rather use an AS number and at least a ipv4/24 and ipv6/48 to
leverage with that, so I don't have to nat the various emulations.
(and funding, anyone got funding?) Or, as the code is GPLv2 licensed,
to see more test designers setup a testbed like this to calibrate
their own stuff.
Presently we're able to test:
the broadband forum udp based test:
There's also a virtual machine setup that we can remotely drive a web
browser from (but I didn't want to nat the results to the world) to
test other web services.
Rpm mailing list
Rpm at lists.bufferbloat.net
Starlink mailing list
Starlink at lists.bufferbloat.net
Starlink mailing list
Starlink at lists.bufferbloat.net
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the LibreQoS