[c-nsp] asymmetrical performance inside core

[Tim Densmore]

On 10/01/2012 10:31 PM, Mikael Abrahamsson wrote:
> On Mon, 1 Oct 2012, Tim Densmore wrote:
>
>> Stab in the dark, but have you verified you don't have a duplex 
>> mismatch anywhere?  IME, they can look exactly like that.
>
> I seriously doubt you'd be able to get 30 megabit/s over somewhere 
> with a duplex mismatch. My testing was done 7-8 years ago so things 
> might have improved regarding TCP congestion algorithms, but I doubt 
> it'd make that much of a difference. My experience is/was that with a 
> duplex mismatch and 10 or 100 megabit, you seldom got more than around 
> 2 megabit/s of effective throughput using TCP.
>
Fair enough.  IME, recently on 100m eth with dup mismatch, I've easily 
been able to pass 30+ mbps one direction, but only 3 - 6 mbps in the 
other while there was live traffic on it - flash test, iperf, udp, tcp, 
etc.  I'm not super hip on duplex mismatches, but my understanding is 
they always tend to affect one direction more than the other, especially 
in scenarios where you have a large push or pull bias on a link.  That 
said, like I said, it was a total stab in the dark since the OP has 
revealed very little about what his network is comprised of, only that 
he had severe speed mismatches showing up via speedtests.  We know it 
contains "radio" but not even the brand.  And "core" + "radio" usually 
leads me to expect mikrotik + ubiquiti these days, not cisco,  but I may 
be spending too much time reading wisp lists.


TD