Comment by wanderingbort | Hacker News Reader

Comment by wanderingbort | Hacker News Reader

Parent Full thread

> Do the actual math.

You first. You made a claim which was woefully inaccurate, I called you out in an admittedly childish way and now the burden of proving you wrong is on me?

Here is my ballpark, Both sides will want redundant fiber connections between two edge routers at 10gbps. SFP+ 10g from cisco is ~60 bucks. We need at least 4 (2 on each side), plus a reserve of 4 for faster recovery from failure. In addition, lets assume that we amortize some cost of the _edge router_ for this service. A refurb 16 port SFP+ module for 10g fiber is around $6000 and we will need 2 of them in 2 separate chassis on EACH SIDE and lease Colo space for it but we will amortize it. So, $6k + $1k for the chassis + $800 for the supervisor card. to lease space for this: $1000k/month per chassis(5u at 200/u in a major market carrier hotel) / 16. We need cabling too but lets assume thats commodity.

up front costs so far:

$60 (sfp modules) * 8 = 480

$7800 edge router * 4 / 16 = 1950

so $2430 for equipement (amortized)

1k/month/chasis = 4k/month / 16 = 250/month recurring

100 customers over 2 years = 1950 + 6000 / 100 / 24 = $3.31/month/customer before we even talk about monitoring/maintenance/legal/etc

I have not even begun to dig into the practical costs, and I'm at 3x your target. Your move.

The ISP doesn't have costs on each side, the other side is the peer's equipment. And you're just intentionally setting money on fire. You can get more than 16 ports into five rack units, which would result in less amortization on the chassis and less colo per port. And 5U worth of colo is not $1000/month almost anywhere, but that was >75% of your total cost.

Which is still using the numbers where equipment only lasts two years and you're expecting 100% continuous usage from all customers. And even using all these ridiculous numbers it's still only $3.31/month per customer. If we adopted a 10:1 oversubscription rate (which is still conservative) it would be $0.33/month even with you setting money on fire, having the ISP pay for both sides, and using full double redundancy rather than some kind of N+1 or N+M.

Peering is just not a major cost for ISPs.

You are moving the goal posts. I put my estimate in the same context you put your estimate in (no over subscription, 100% continuous usage etc).

The ISP doesn't pay both sides, but someone pays somewhere. So, if we are talking about the "cost of peering" it doesn't serve to talk about it from only one side.

You can get more than 16 ports in the 5u unit, so rent is probably under-amortized but not by the two order of magnitudes that would make up the difference between my estimate and your original.

Peering is not a major cost for ISPs because ISPs don't have a lot of peers. As more services want direct peering relationships, there are some overhead costs which go up to manage the explosion of peering relationships.

My initial objection was that you over-simplified. I feel pretty fine resting on my argument as is. You may have the last word. I cede all my remaining time to whomever wants it.

> The ISP doesn't pay both sides, but someone pays somewhere. So, if we are talking about the "cost of peering" it doesn't serve to talk about it from only one side.

But we're talking about the cost of peering paid by the ISP, because the ISP doesn't have to cover costs paid by somebody else and no one else is objecting to paying their part of it.

> You can get more than 16 ports in the 5u unit, so rent is probably under-amortized but not by the two order of magnitudes that would make up the difference between my estimate and your original.

Juniper MX240 supports 80 10GbE ports in 5U. Cisco's product search is lame but I expect they have something similar too.

> Peering is not a major cost for ISPs because ISPs don't have a lot of peers. As more services want direct peering relationships, there are some overhead costs which go up to manage the explosion of peering relationships.

That isn't necessarily true either. Only the major transit providers and CDNs like Level 3 and Akamai and the major players like Google that are so big they're effectively their own transit provider/CDN have any interesting peering.

All the smaller players get connected through one of the transit providers, because Joe's Website only wants to hook up from wherever Joe's servers are, not at every carrier hotel in the world like Google does. So they pay a transit provider to handle that and the ISP only has to deal with the transit provider.

The problem comes when the ISPs want to charge major transit providers monopoly rents for peering (which they would have to pass on), or refuse to upgrade the links to them when they're saturated in order to force their big customers (like Netflix at one point) to pay monopoly rents for peering directly.

> My initial objection was that you over-simplified.

It's supposed to be a simplification. The point isn't to make a full accounting, it's to compare orders of magnitude. Even using your numbers with unrealistically conservative assumptions about everything, they're still a single digit percentage of the typical customer's bill.

Reply on news.ycombinator.com