Do ATM-based Exchange Points make sense anymore?

Thu Aug 15 15:04:18 UTC 2002

Hi all -

I have walked about 30 people through the "Do ATM-based Internet Exchange 
Points make sense anymore?" white paper and have received some really good 
feedback, suggestions and price points to calibrate the Peering Financial 
Model. I have applied these calibrations and I am ready to release the 
paper for wider review, but I'd like to share first the assumptions and 
calibration points for the model along with a few of the more interesting 
observations.

The Business Case for Peering at an ATM-based Internet Exchange Point 
Peering looks pretty dismal in todays market.  As I mentioned in an earlier 
message, the dominant issue is that transit and transport have dropped 
dramatically,while the cost of ATM-based peering has not dropped in kind. 
In todays market (from quotes shared with me) we see:

Assumptions and Calibration Points
------------------------------------------------------
Transit $125/Mbps with 500Mbps commit, $100/Mbps with 1000 Mbps commit.
Transport (DC-ASH) $2500/mo for OC-3, $5000/mo for OC-12
Eth-IX fees: $2500/mo for 1/2 rack and FastE, $5000 for 1/2 rack and GigE
Eth Framing Overhead: 6%
HDLC Overhead: 4%
ATM-IX fees: $11,000/mo for OC-3, and $26,000 for OC-12 transport and Port
ATM cell tax: %20
Effective Peering Bandwidth=75% average utilization of available bandwidth
(this means we assume that ISPs (for policy reasons) upgrade the peering 
infrastructure when the average utilization is 75%)

These numbers are empirical and based on averages from the Internet 
Operations Community. The paper footnotes the sources.

Observations
------------------------
When these numbers are plugged into the Peering Financial Models, we see 
that OC-3 ATM-based peering is "Effective" (less expensive than transit) 
for the very narrow range of 88Mbps-90Mbps. If an ISP can't send at least 
88Mbps over the OC-3 to the ATM-IX, it would save money by simply buying 
transit. At 90 Mbps the OC-3 ATM must be upgraded. This narrow range leads 
me to believe that OC-3 ATM peering is simply not cost effective. Under the 
same assumptions (OC-3 into  FastE IX), the Fast Ethernet-based Effective 
Peering Range is  40Mbps-70Mbps, a more reasonable range for medium scale 
peering.

Applying the model to the ATM-OC-12 we see the Peering Breakeven Point is 
260Mbps; if you don't send at least 260Mbps to the peering population then 
you should prefer simply to purchase transit. This peering infrastructure 
scales to 375 Mbps at which time it must be upgraded. In this Effective 
Peering Range the cost of traffic exchange ranges from $100/Mbps down to 
$69/Mbps when the Effective Peering Bandwidth is fully utilized.

The same analysis applied to Gigabit Ethernet shows a much lower Peering 
Breakeven Point (100Mbps) with a broader range, scaling up to 448Mbps 
before the OC-12 must be upgraded, at which point the cost of peering 
traffic exchange is $22/Mbps.

The bottom line is that the cost of the ATM Peering infrastructure, and the 
dropping price of transit and transport, have conspired to destroy the 
value proposition of ATM-based Internet Exchanges. Ethernet-based IXes are 
less expensive and have a broader "useful life", defined in this paper as 
"Effective Peering Range."

As I walked folks through this paper I got the sense that most folks had 
not done this  analysis and we opened some eyes here. Thanks to those who 
provided the empirical HDLC, ATM, and ethernet overhead figures.  Including 
these provides a more fair comparison between ATM and Ethernet-based IXes.

If you would like a copy of the paper please send e-mail to wbn at equinix.com 
and I'd be glad to send you a copy. As always, I'd love to hear your 
feedback; that is how these papers become valuable resources for the community.

Thanks -

Bill