[c-nsp] asymmetrical performance inside core

[Phil Mayers]

On 10/02/2012 12:03 AM, Eric A Louie wrote:
> I'm inheriting a problem that I could use some ideas to troubleshoot.
>
> Speedtesting from within my core to other locations within the core give me
> "asymmetrical" performance. I've traced the routes and the path is the same
> download and upload, but I'll get very good download speeds (30Mbps) and
> terrible upload speeds (3Mbps, 6Mbps)
>

Good compared to what? You haven't really given us any idea of what kind 
of network it is, and what you're expecting - and what that's reasonable.

> I've looked at MTU and it doesn't appear that there's a lot of packet
> fragmentation/reassembly

There really shouldn't be *any*.

> I've looked at total bandwidth utilization during the test times across the
> backhaul and I'm not saturating the bandwidth in either direction
>
> I've looked at the interfaces through which the packets are traveling, and there
> are no CRC errors, no drops, no retransmissions
>
> The physical layer radio transmissions seem to be good both directions with the
> performance testing at that level.

Ugh. Radio.

This kind of problem is super-hard to troubleshoot on any network. My 
usual technique is to jump straight to getting a tcpdump of the packets 
at ingress to the network, and egress to the machine, and using 
wireshark or (preferably) tcptrace to analyse them. If you're getting 
slow downloads and have respectable loss and RTT in ping, odds are the 
congestion control algorithms think there's a problem.

iperf and such like are your friend in this situation. But you'll need a 
known-good machine upstream to source/sink traffic from/to.