Stephen Wilcox Telecomplete Ltd. Rutherford House Lloyd Street North Manchester M15 6SE. UK Peering Economics of a Long Distance ISP Operator October 2004. Draft 03 1. Scope of document This paper takes the point of view of an operator having long distance or international network who I will call BigNet. BigNet will be evaluating the benefit of establishing peering in a particular location with another operator who does not have similar network coverage and hence will only be interested in peering at a single location, I will refer to them as SmallNet. The question that arises is whether for BigNet to peer with SmallNet increases or decreases profits for BigNet. 2. Assumptions Some of these exist to reduce complication in the modeling, some are to allow focus on only the economic aspect: All operators are ISPs with their own ASNs giving consistent routing announcements Hardware and maintenance costs do not change by peering No intervention from sales (not peering with potential customers) We are ignoring performance benefits (from reduced latency, congestion etc) BigNet currently sees all of the Internet through non-transit paths The author's monetary figures are reasonable BGP routing is hot potato (i.e. shortest exit) BigNet has a fully meshed network of nodes with equal cost transports BigNet is predominantly a transit network to other networks We're not concerned with ratio requirements for these or other existing peers 3. Network and Financial Data We need to determine the costs involved in this model, taking figures based on "A Business Case for Peering in 2004" we will use the following: Peering costs are $25/Mbps/month Transport into a peering location costs $10/Mbps/month Average sale price of transit is $50/Mbps/month These prices are bidirectional therefore when considering unidirectional traffic flows we need to use these figures halved i.e. $12.50/Mbps and $5/Mbps, we will generally do this when looking at peering traffic but not transit traffic. We will assume BigNet is present in four nodes each with peering capabilities and that BigNet peers with SmallNet's transit provider in all four nodes. SmallNet is only present and able to peer at one node. 4. Scenarios 4.1 Non-peering scenario If BigNet does not peer with this SmallNet it will exchange traffic to SmallNet through SmallNet's transit provider via the four peering nodes. As routing is hot potato this will mean that inbound traffic to BigNet is coming exclusively via the one node where SmallNet is present (as SmallNet takes transit at this node and the transit provider will hot potato the traffic straight onto BigNet). Also by hot potato, outbound traffic will exit immediately onto SmallNet's transit provider's network. The economic implications of this for BigNet are that outbound traffic incurs only the cost of peering whereas inbound traffic will incur the cost of peering plus transport costs. Therefore: outbound traffic costs BigNet $12.50 = $12.50/Mbps inbound traffic costs BigNet $12.50+$5.00 = $17.50/Mbps overall traffic costs for BigNet: = $30.00/Mbps 4.2 Peering scenario If BigNet peers with SmallNet at a single node we now have adjacency of as paths and all traffic will come in and out via that one node, in economic terms outbound traffic incurs cost of peering plus transport costs and inbound traffic incurs cost of peering plus transport costs also. Therefore: outbound traffic costs BigNet: $12.50+$5.00 = $17.50/Mbps inbound traffic costs BigNet: $12.50+$5.00 = $17.50/Mbps overall traffic costs for BigNet: = $35.00/Mbps 4.3 The as-path effect In the non-peering scenario traffic to/from BigNet's customers is being carried to SmallNet via BigNet and other transit providers. The choice over whether BigNet or another of the transit providers is carrying the traffic will be a result of the BGP path selection algorithm which is likely largely dependent on the as paths available. In the non-peering scenario paths for typical customers would be like: BIGNET_TRANA_SMALLNET TRANB_TRANA_SMALLNET Traffic would travel via either of these two paths (for an average sample of customers this will likely be an even split). In fact customers may have more than two transit providers so rather than BigNet taking 50% of the available traffic the percentage is probably lower. To compound this it is also possible that one of these alternative transit providers to our customers is peering already with SmallNet in which case BigNet would have a longer as path and not be chosen to carry any traffic for such a customer. When we move to the peering scenario the above paths would become: BIGNET_SMALLNET TRANB_TRANA_SMALLNET Now BigNet will be preferred over the alternative transit resulting in all traffic being carried by BigNet. Other scenarios are available, such as where an alternative transit provider already peers with SmallNet and hence BigNet does not carry any traffic unless it also peers directly, or for customers of customers where as path is longer and more complicated. The detail is unimportant, the principle however is that in reducing the as path we now become chosen to carry traffic much more than we were previously. 5. Peering Decision 5.1 Method We can now construct a simple economic equation as follows: profit = (Number of Mbps from transit customers to SmallNet [peering/non-peering]) x ( (charge per Mbps to transit customers) - (cost per Mbps [peering/non-peering]) ) And by working out the profit in the non-peering scenario and the peering scenario we can easily see which of the two scenarios makes BigNet the most money and hence what the correct decision would be to maximize revenue. But, although the equation is straight forward the variables going into it are not and this is crucial to the decision. Some of the issues include: o Measurement of traffic with SmallNet is simple when the peering is in place but before the peering is established traffic is passing via other networks and many networks have difficulty measuring traffic to a non-adjacent network. o Charge per Mbps to transit customers will vary, an average figure may be used but depending upon the nature of our prospective peer (SmallNet) this may skew the average. o Our transport costs are simple, traffic in reality may show regional trends giving non-average costs o Because traffic is two way but we only tend to be interested in the higher of the in and out values figures do not add quite linearly, we get some interesting effects from aggregation which usually results in a perception of having less coming out than what is being put in. o Also because of two way traffic our model may break down for non-equal ratio scenarios where we may have large flows in one direction but are only considering our income from static 1:1 customers. We cannot apply the ratio to customers as they are independent of the peers' ratios. With these caveats in mind, using some sample figures we can look at how our model reacts to changes in our variables. 5.2 Varying traffic increase from peering Fixing the following parameters: Ratio=1:1, Transit=$50/Mbps TRAFFIC --NON-PEERING-- ----PEERING---- PROFIT GROWTH COSTS REVENUE COSTS REVENUE INCREASE 10% $1500 $2500 $1925 $2750 $-175 30% $1500 $2500 $2275 $3250 $ -25 50% $1500 $2500 $2625 $3750 $ 125 70% $1500 $2500 $2975 $4250 $ 275 90% $1500 $2500 $3325 $4750 $ 425 As we can see for small traffic growth we get a reduction in profit but for any value at or above 33.3% in this model we get an overall increase in profit which is linear (proportional to the traffic growth). 5.3 Varying the average price of transit Fixing the following parameters: Ratio=1:1, Traffic Increase = 50% TRANSIT --NON-PEERING-- ----PEERING---- PROFIT PRICE COSTS REVENUE COSTS REVENUE INCREASE $ 25 $1500 $1250 $2625 $1875 $-500 $ 50 $1500 $2500 $2625 $3750 $ 125 $ 75 $1500 $3750 $2625 $5625 $ 750 $100 $1500 $5000 $2625 $7500 $1375 $125 $1500 $6250 $2625 $9375 $2000 Here we see that if the average transit sale price is too low we would again make losses with the break even point being at $45 in this model. However as an overall average we would expect this value to be somewhat higher. The relationship between transit price and increase in profit is again linear and its interesting that for the range of values which we can expect to see in the market this increase is quite sensitive. 5.4 Varying traffic ratio of the peer Fixing the following parameters: Transit=$50/Mbps, Traffic Increase = 50%. Note this is the ratio as viewed from BigNet. TRAFFIC --NON-PEERING-- ----PEERING---- PROFIT RATIO(I/O) COSTS REVENUE COSTS REVENUE INCREASE 1:5 $1333 $2500 $2625 $3750 $ -42 1:2 $1417 $2500 $2625 $3750 $ 42 1:1 $1500 $2500 $2625 $3750 $ 125 2:1 $1583 $2500 $2625 $3750 $ 208 5:1 $1667 $2500 $2625 $3750 $ 292 Again we have a range of profit increases, although it is not clear from the way we normally write ratios this is actually another linear relationship with the break even point occurring for a ratio of 1:3. This produces a somewhat counterintuitive result that peering with content heavy networks increases profit more than peering with eyeball heavy networks. Looking at the math behind the above numbers the reason is because in the eyeball heavy ratio we are actually looking for a profit -increase- and in the non-peering scenario we are already making a large margin (relative to other scenarios presented here) and peering doesn't create enough extra margin. Now considering the real world what we are actually seeing here for the content heavy network is that because of hot potato routing we are already carrying this traffic near to the source because the content network's transit provider will be immediately offloading it to BigNet so by peering we do not actually change how we carry the traffic but we do see an increase in the volume and hence a profit increase. Considering further implications of the above conclusion is beyond the scope of this document but it is logical that for inbound traffic to BigNet peering and transport costs play a major role and our model is perhaps not sufficiently complicated enough to adequately report that. If transport/peering costs to a distant node were higher than the transit sale price we would have a scenario where all traffic is loss making and we might actually want to avoid carrying that traffic wherever possible and not peering is certainly one way if we can shunt that traffic to one of BigNet's competitors! 5.5 Results from variations We can see that the three variables we have played with above all have an effect on how much profit is increased by so these are important to any economic decision on peering. Within the bounds of what we would expect to see in the real world it would appear that all of these can easily swing the decision to showing an increase in profit. It is difficult to say which are the most important as this depends upon the range of values for a particular operator which may vary considerably however for the values we have input into the model here it would seem that the order of importance would be transit cost then percentage traffic increase and finally traffic ratio. One thing in this model which is notably different from real world peering decisions is that we look at the percentage increase of traffic rather than absolute volume. As to which is the correct approach is a decision left to the operator but it is obvious that for small volumes of traffic other factors could become major which we have not modeled here such as the overheads associated with managing a new peer, possible interface and cross connect costs and the load of a bgp session on a router. Indeed other costs may play a major part in the peering decision for other reasons too, such as for our figure of transit cost, if the overheads of managing a transit customer, paying sales commissions etc are high they may produce a net transit cost figure that is somewhat lower than we have used here. One final thing to take into account in a more complicated model is some of the variables which we have kept fixed here. These would be the costs of transport circuits which may vary a lot even within an organization operating in different global regions. Also some of our assumptions on the bidirectional flow of traffic are simplistic: we are not incorporating the aggregates that occur which appear to see a reduction in traffic when we combine many small flows into a few large ones; for increasingly non-even ratios the larger direction becomes dominant and we can't assume a 50% cost for this one direction; we have also used strictly unidirectional figures for our peering costs but bidirectional figures for our transit revenues. 6. Summary By this point in the document most readers will already have a number of complaints as to how the model is deficient and how some of the assumptions made here are difficult to reason with but this paper was not written to be fully comprehensive. As we can see above there are a lot of variables and unknown factors which different operators may have their own values for and views on which work for their own unique scenarios. What we do have however is a basic framework of how peering works at a certain level and from this we have seen a number of critical factors that need to be evaluated when looking at any network build and peering decision. Both small and large operators can use this information to understand how peering may or may not benefit them. References William B Norton, Equinix - A Business Case for Peering in 2004 (2004) Steve Gibbard, PCH - Economics of Peering (2004) Acknowledgments The following people contributed ideas or reviewed this draft especially Peter Cohen, Steve Gibbard, Sylvie Laperriere and William Norton. Thank-you to everyone who helped! Copyright This document is copyright (c) Stephen Wilcox 2004. All rights reserved. Permission is granted to copy and distribute the text of this document providing it is complete and unedited in content. Disclaimer This document and information contained herein are provided on an "as is" basis. The author accepts no liability for discrepancies caused by the assumptions, source data or generality of this approach.