[Cake] fq_codel on 3g network in Mauritius
moeller0 at gmx.de
Tue Jul 19 06:16:27 EDT 2016
> On Jul 19, 2016, at 10:09 , Jonathan Morton <chromatix99 at gmail.com> wrote:
>> On 19 Jul, 2016, at 07:29, Loganaden Velvindron <loganaden at gmail.com> wrote:
>> I've been playing with fq_codel on 3g internet connection. the 3g
>> internet box has no bridge mode, and just forwards every packet to the
>> Openwrt router.
>> Here are the results without:
>> And here is the result with fq_codel on and 300 ms target latency:
>> Is there anything that could be done to get the rating up to A+ by
>> tweaking the code ?
> You say fq_codel, rather than Cake. Presumably it is paired with some sort of shaper, such as HTB? This is important, because I think the shaper is influencing part of your results.
> On downstream, you have one high latency sample (750ms) in the middle of a series of reasonable ones (250ms). This implies a momentary glitch in your connection, which isn’t unusual with wireless links. Re-measuring might eliminate it.
> On upstream, you have two very high latency samples at the *beginning* of the run, which then clear out to approximately the baseline latency. This is a classic sign that your shaper is letting a burst of traffic through before actually starting to control it, which is typical behaviour for token-bucket shapers. That initial burst collects in the dumb queue of your 3G modem and takes time to drain away.
> Cake uses a shaper carefully designed to *not* burst in that manner, while still maintaining full throughput regardless of timer resolution and latency. Using Cake instead of fq_codel+HTB will therefore probably improve your upload characteristics.
I believe that the way sqm-scripts simple.qos and simplest.qos use the HTB shaper are not subject to this specific issue (we do not configure/allow HTB to burst at all). Looking at a number of my speed tests seems to support this observation (https://www.dslreports.com/speedtest/636817 simplest.qos, HTB+fq_codel no initial bursting).
> - Jonathan Morton
> Cake mailing list
> Cake at lists.bufferbloat.net
More information about the Cake