Peering and Network Cost

Max Tulyev maxtul at
Tue Apr 21 17:37:02 UTC 2015

That's generally good idea, but average TCP session speed depends not
only your side of connection, but another side as well.

On 18.04.15 07:58, Mark Tinka wrote:
> On 17/Apr/15 15:05, Max Tulyev wrote:
>> One more interesting thing.
>> If you buy IP transit, mostly you are paying by exact bandwidth, per
>> megabit. If you buy IX peering port, you are paying for port. This means
>> Tranist ports are overloaded or close to it, while IX ports usually
>> always have some extra free capacity.
>> In practice, this mean if your customer download some file using IX way,
>> speed will be much higher that same file reachable by IP transit.
> This depends entirely on how you run your network. If you run links hot,
> you can't guarantee anything (keeping in mind that your less congested
> exchange point ports does not mean other exchange point members are in
> the same position also).
> We, for example, buy transit or peer with a minimum of 10Gbps port, with
> the ability to push traffic at line rate if needed. We do not allow
> ports to run hot (typically upgrading them anywhere from between 50% -
> 70% utilization). I appreciate that not everyone can be in this
> position, while others can be even more aggressive with their
> "over-engineering", but this kind of information is hard to quantify
> reliably.
> There is also backhaul from the interconnect point into the backbone to
> think about, but that follows a similar strategy.
> Mark.

More information about the NANOG mailing list