Merit Web

[To print this page click in this portion of the window before you hit your print key!]
Back to: NANOG Home

QoS and Ebone: Issues and Potential Solutions

                Presentation Notes by Sean Doran


Ebone, the organization that provides global Internet connectivity in Europe and the U.S., has been testing Cisco's RED (Random Early Detection) implementation. The code has some frailties, is difficult to deploy on non-VIP (Versatile Inter

face Processor) interfaces, and the "knobs" seem insufficient. However, RED clearly works to improve customer line utilization. We therefore want to deploy it on all lines that see non-zero average queue lengths and drops.

Ebone's engineering policy is to ensure that bandwidth contention in normal circumstances only happens at the edges (i.e., the customer connections all can be full and the average queue lengths in the networks will be very near zero).

Ebone is therefore QoS-transparent in normal circumstances - but what about abnormal circumstances? We obviously try to minimize these through overcapacity and redundant topology, but things do tend to fail. When they fail, there is a chance that aver age queue lengths will grow until RED can shrink them. However, at this point QoS becomes practicable for the duration of the fault.

Does practicable mean useful? We don't know yet, and it is probably more useful for us to solve a bigger problem.

Pricing Problems

Ebone's biggest service problem is not quality (we do a good job) but price: we are hideously expensive (about 200k$/1Mbps/year).

The price is the same in all locations. This favours customers connecting to smaller sites, such as Bratislava, but disfavours customers clustered together in large locations.

Ebone is in effect owned by its customers and is not-for-profit; it is easy to become an Ebone customer and easy to walk away. Multihoming is expected. Customers clustered together in large locations are therefore likely to colocate around exchange p oints and interconnect directly, using Ebone for long-distance connectivity (to places like Bratislava). This, of course, can be inefficient.

If instead a customer can connect directly to Ebone and pay a price comparable to IX fees and capital equipment costs for purely local connectivity, and our normal (expensive) cost-based price for international connectivity, this is a good thing.

There is a price-perversity at work in Europe; U.S. capacity is cheaper from some locations than European capacity. This is particularly true in the UK and the west coast of Europe. Consequently, most Ebone customers have their own lines of various s izes (ranging from 155Mbps to 2Mbps or smaller) to U.S. providers.

Ebone does provide U.S. capacity; it is cheaper to offer in several locations than European capacity is. The pricing scheme has been based on an arrangement that has a customer take 2/3 of its capacity from U.S.-based sources, and 1/3 from European-ba sed sources, and leads to a percentage discount.

It would be more flexible to use the same mechanism that allows local and long-distance traffic to be separated to further divide long-distance traffic into European and U.S. categories. Also, it would be competitive price-wise with the thin circuits across the ocean.

What’s the Solution?

There are two approaches; one is more radical economically, the other is more radical technically, and relies on QoS-like mechanisms.

PROPOSAL #1: Use separate rate-limited/rate-controlled queues. The goal is a bill with three separate categories (local, long-distance (Europe, transatlantic) each with its own flat-rate price.

- Inbound from customer is easy: we use BGP communities to tag routes at ingress, and can use those communities in building FIBs that select different rate-limits per community.

- Customer traffic towards a prefix we hear from the U.S. will go out one rate-limit.

                - Traffic towards a prefix we hear locally will go out another, etc. - Outbound is harder. one approach is to do an RPF-like check as a classifier; we sort traffic based on the community of the entry in the BGP RIB of the source address. This is straightforward technologically, but has flaws in the face of asymmet rical traffic

- The other approach is to use diff-serv, which somewhat liberates the TOS byte in IPv4 headers. Mark each packet at ingress with a description of where the ingress is and use those marks to sort traffic into rate-limit queues.

- This solves the asymmetry problem and also gives flexibility -- we can make certain traffic more droppable, so when there are abnormal circumstances "expensive" traffic will take precedence over less-expensive traffic (WRED). The problem: markin g packets at ingress is expensive, and has a scalability problem. There are only so many codepoints in diff-serv.

- The final approach (in some future world): use MPLS. Diff-serv, after all, is a breeding ground for MPLS ideas.

- [plug: we will also want a fourth queue for traffic that remains within North America]

PROPOSAL #2: packet-mile billing (modified Lothberg idea)

                - On each VIP maintain an accurate AS*AS traffic matrix

- Offline publish "long-term MEDs" that associate a cost from a particular interface to any given AS in normal circumstances. (Normal MED becomes interesting too ...)                 - Generate a bill: #-of-packets-to-ASn * cost-to-ASn - This can be seen as a massive extension of the multi-queue system in proposal #1 combined with dynamic setting of the per-queue rate limit. We can combine this with flat rate traffic that is more droppable than the per-packet-charged traffic. Incremental ‘outside-in’ deployment does indeed work.

- This solution is scalable, not hard on routers, and has some useful side effects (encourages bigger and fewer packets, encourages content replication, etc.) The flaws: requires accurate counting/billing ceilings. Other flexibility packages are hard to maintain. We have the "pandora's box" attitudes to settlements.

Can we move from Proposal #1 to Proposal #2? We don’t know, but Peter and I will certainly experiment. Both proposals make peering decisions somewhat easier; purely local peering that doesn't use international lines at all will be very inexpensive, and there is a negotiation framework for more extensive peering.