I don't read this list continually, but do archive it; your note was
flagged for me to comment on.
This is probably for Jim Gettys directly, but I’m sure most others have
input. I could of sworn that that there was some test made to detect it
directly on switches and routers? Sort of like iperf, but to test
bufferbloat specifically given the OS stack which is going to have issues
as well, as shown on bufferbloat.net <http://bufferbloat.net/>.
We recommend Toke Høiland-Jørgensen's
"flent"
https://flent.org/ for testing connections/devices/gear. It uses "netperf"
transfers to load the link (by default with 4 simultaneous TCP connections
in both directions, IIRC), and then runs another test (by default "ping")
at the same time to test the connection under load.
Turning on a netperf server is just as easy as turning on an iperf server
(and the results are better, and netperf's maintainer responsive).
See the documentation/paper on Toke's web site. The "RRUL" test
("Real-Time Response Under Load") is the one we use most/is best shaken
down. I'm sure Toke would love help with other tests.
Gives you lots of useful graphs, will do diffserv marking, etc...
>
> On 7/21/16, 2:19 PM, "NANOG on behalf of Jay R. Ashworth" <
>
>
>
>>> From: "Janusz Jezowicz" <janusz@speedchecker.xyz>
>>
>>> Since this morning Speedtest.net is not accessible in Chrome
>>> Reason:
>>>
https://www.google.com/transparencyreport/safebrowsing/diagnostic/#url=c.speedtest.net
>>>
>>> For any ISPs/content providers linking to speedtest.net you may want
to
>>> swap links to a different website or host your own speed test.
>>
>> So far, I am very pleased with how it works, though I think it's letter
>> grades on speed are a bit pessimistic (65Mbps is a "C").
Most applications are as sensitive/more sensitive to latency than to
bandwidth
; see the research in the field, for example, for web browsing. For web
browsing, you are at the point of diminishing returns on bandwidth after a
few megabits/second, for most use
.
For telephony, the metric is always the lower the better, and not more
than 100ms or so (continental delay).
So it is entirely appropriate in my view to give even "high speed"
connections low grades; it's telling you that they suck under load
, like when your kid is downloading a video (or uploading one for their
friends); your performance (e.g. web surfing) can go to hell in a
hand-basket despite having a lot of bandwidth on the
connection. For most use, I'll take a 20Mbps link without bloat to a
200Mbps one with a half second of bloat any
day.
It will work reliably, I'll be able to make my phone calls without
problems, I'll be able to frag my friends with the best of them, etc...
Even video playback gets wonky with bad bufferbloat: the player's control
loop is interacting with the (wildly excessive due to bloat) TCP control
loop and can't find a good playback point; seeking also becomes slow, etc.
Activities such as web browsing can/does cause transient latency on a link,
since most links are not doing decent scheduling; the damage is done
anytime the link gets used by anyone, for anything, including web surfing
as well as background activities such as backup or system update.
So no, I don't think dslreports grades pessimistically: it's just that bad
bufferbloat is so *blinking* common and bad. And I had nothing to do with
setting the scoring system: that's the opinion of the dslreports test's
author; but I think Justin has done a good job choosing the grades to boil
down the quality of a connection to something mere mortals (your
customer's) will understand. So my hat is off to Justin for doing a great
job.
