[Netalyzr] netalyzr, java, bufferbloat and 50Mb/s links

Jim Gettys jg at freedesktop.org
Mon Apr 16 08:33:42 PDT 2012

On 04/16/2012 11:21 AM, Nicholas Weaver wrote:
> On Apr 16, 2012, at 8:10 AM, Maciej Soltysiak wrote:
>> Hello!
>> As part of my regular routing related to playing with home router firmware I noticed that recently netalyzr started showing low latency and buffering, even on stock router firmware.
>> I've just read "Netalyzr: illuminating the edge network" and linked 2 facts together:
>> a) I've been upgraded to 50Mb/s link (from 6Mb/s)
>> b) the whitepaper says java puts a cap to 20Mb/s
>> So I guess netalyzr tests are not able to fill my cable modem buffers.
>> Is there a way to run the network latency/buffering test so that it is not hampered by java?
> Not our test, but there is one that can work:
> Do a background ping of www.google.com
> During the ping, run Ookla's speedtest.net speedtester to load the network, and watch how the pings change.
> The odds are, with a 50 Mb/s link, you don't have much of a buffer problem since most systems have the critical buffers sized in capacity, rather than delay, so by upping your bandwidth by 10x, you effectively make the buffer 10x shorter in time.
On a 50Mbps cable link, I still have hundreds of milliseconds of
buffering (400ms, IIRC; right now I have taken other measures to solve
that problem on my own home network).

The kiss of death is when you take modern equipment but then run it at
low speed (e.g. a DOCSIS 3 modem, but only buy 10Mbps of bandwidth). 
Then you end up with truly monumental buffers (or you sometimes do
anyway, as there was no thought whatsoever about how much buffer to
use.  So my modem, when run at 20Mbps, was around 1.2 seconds, as you
might expect.
                    - Jim

More information about the Netalyzr mailing list