Question on the topology of Internet Exchange Points
A typical Internet Exchange Point (IXP) consists of one or more network switches <http://en.wikipedia.org/wiki/Network_switch>, to which each of the participating ISPs connect. We call it the exchange-based topology. My question is if some current IXPs use directly-connected topology, in which ISPs just connect to each other by direct link, not through a network switch?? If so, what's the percentage of this directly-connected case? Kai
On Thu, Feb 14, 2008 at 11:02:54AM -0600, Kai Chen wrote:
A typical Internet Exchange Point (IXP) consists of one or more network switches <http://en.wikipedia.org/wiki/Network_switch>, to which each of the participating ISPs connect. We call it the exchange-based topology. My question is if some current IXPs use directly-connected topology, in which ISPs just connect to each other by direct link, not through a network switch?? If so, what's the percentage of this directly-connected case?
Kai
the "directly-connected" case - over point2point link is not per se, an Internet Exchange Point (IXP) in that there is no chance of multiplexing the link to connect more than one provider over that direct link. the direct link can be a dedicated fiber pair, a cat5 cable, conditioned copper pair or coax or combination of these layer one transmission media (yeah, sat, microwave, avian carrier etc...) depending on proximity and cost. latency is usually less of an issue here, as is buffering, since there is a single endpoint. Its also much easier to maintain security associations on direct links. --bill
kch670@eecs.northwestern.edu ("Kai Chen") writes:
A typical Internet Exchange Point (IXP) consists of one or more network switches <http://en.wikipedia.org/wiki/Network_switch>, to which each of the participating ISPs connect. We call it the exchange-based topology. My question is if some current IXPs use directly-connected topology, in which ISPs just connect to each other by direct link, not through a network switch?? If so, what's the percentage of this directly-connected case?
when i last worked at PAIX, the private interconnects wildly outnumbered the switch connection. the model seemed to be, use the switch to reach all of the other participants, but whenever you had a hot neighbor, get a PNI. in other words there appeared to be no "exchange-based topology", more like a "hybrid exchange and PNI topology." -- Paul Vixie
On Feb 14, 2008, at 7:06 PM, Paul Vixie wrote:
kch670@eecs.northwestern.edu ("Kai Chen") writes:
A typical Internet Exchange Point (IXP) consists of one or more network switches <http://en.wikipedia.org/wiki/Network_switch>, to which each of the participating ISPs connect. We call it the exchange-based topology. My question is if some current IXPs use directly-connected topology, in which ISPs just connect to each other by direct link, not through a network switch?? If so, what's the percentage of this directly- connected case?
when i last worked at PAIX, the private interconnects wildly outnumbered the switch connection. the model seemed to be, use the switch to reach all of the other participants, but whenever you had a hot neighbor, get a PNI.
in other words there appeared to be no "exchange-based topology", more like a "hybrid exchange and PNI topology." -- Paul Vixie
Obvious as it is, if one of your peerings on an IX gets big in terms of in/out volumes, you HAVE to secure it by PNI. You need a way to prevent the IX's equipments from being a SPoFs between you and that peer. I'm not saying one should convert every single IX peering into a PNI, as I feel both are pretty much required: your smallest peers shall be secured on as many IXes as possible, your biggest ones via PNI. IX peering is mandatory to keep internet routing diversity up to par - and enable small ASes to grow. Also, it is a wrong assumption to state that IX will make you spare money on transit, from my perspective they should be seen as securing multiple narrower paths to the internet. Greg VILLAIN Independant Network & Telco Architecture Consultant
On Feb 15, 2008, at 8:04 AM, Greg VILLAIN wrote:
Obvious as it is, if one of your peerings on an IX gets big in terms of in/out volumes, you HAVE to secure it by PNI. You need a way to prevent the IX's equipments from being a SPoFs between you and that peer.
"HAVE to" is such a strong phrase. First, who said the switch is a SPoF? And since when is a PNI not a SPoF? If the peer is that big, you should peer in more than one place. For instance, LINX has two LANs, or you can use PAIX and Equinix. Connecting to a "big peer" in a single location, whether over PNI or shared switch, creates a SPoF. Peering in multiple locations removes the SPoF, regardless of the method.
I'm not saying one should convert every single IX peering into a PNI, as I feel both are pretty much required: your smallest peers shall be secured on as many IXes as possible, your biggest ones via PNI. IX peering is mandatory to keep internet routing diversity up to par - and enable small ASes to grow.
Using shared for small peers and direct for big peers is a time honored practice on the Internet. But you can justify this in finance, not just engineering. A fiber x-conn costs less than an IX port (usually). Any peer big enough to take up a significant fraction of the IX port probably justifies the CapEx for a dedicated router port. Does this make things more reliable? Many would argue it does. I would argue that large IXes have amazing uptime these days. The MAEs & GigaSwitches are long gone, public IXes are no longer guaranteed to give you problems.
Also, it is a wrong assumption to state that IX will make you spare money on transit, from my perspective they should be seen as securing multiple narrower paths to the internet.
Do you mean "save money on transit" when you say "make you spare money on transit"? Just want to make sure we are on the same page. That is not an assumption, it is a provable - or disprovable! - fact. If you run the numbers and the IX saves you money, well, it saves you money. If it does not, it does not. Where does the word "assumption" come in? That doesn't mean they are not also additional vectors. But Item #1 does not conflict with Item #2. -- TTFN, patrick
Greg VILLAIN wrote:
I'm not saying one should convert every single IX peering into a PNI, as I feel both are pretty much required: your smallest peers shall be secured on as many IXes as possible, your biggest ones via PNI. IX peering is mandatory to keep internet routing diversity up to par - and enable small ASes to grow.
The converse can also be true - we have a number of members who use the IX fabric as a backup to their PIs with larger peering partners. If you lose a PI carrying a GE of traffic, where does that traffic go? -- Will Hargrave will@lonap.net Technical Director LONAP Ltd
On 14 Feb 2008, at 17:02, Kai Chen wrote:
A typical Internet Exchange Point (IXP) consists of one or more network switches, to which each of the participating ISPs connect. We call it the exchange-based topology. My question is if some current IXPs use directly-connected topology, in which ISPs just connect to each other by direct link, not through a network switch?? If so, what's the percentage of this directly-connected case?
ISPs use a "direct link" (PNI) all the time to peer, and they don't need to be a member/customer of an internet exchange point to do so. In fact, the network you want to peer with might not be available at your local IXP even though they're in your local carrier hotel - then it becomes pretty much the only way to peer. In London, the LINX offer switched *and* unswitched connectivity between members - you can rent fibers from them in order to perform PNI with other members. That the exchange offer this is unusual, and it's offered as an additional service, in order to smooth the process of organising interconnection between member organisations. [www.linx.net ] We (LONAP) don't offer PNI services, and only offer the conventional switch ports, which members normally want so that they can get access to our peering lan and swap some traffic. [www.lonap.net] All exchanges offer this connectivity model. We offer private CUG and member-to-member private VLANs, which is similar, but still passes through the switch fabric. Best wishes Andy
On Feb 14, 2008, at 3:44 PM, Andy Davidson wrote:
On 14 Feb 2008, at 17:02, Kai Chen wrote:
A typical Internet Exchange Point (IXP) consists of one or more network switches, to which each of the participating ISPs connect. We call it the exchange-based topology. My question is if some current IXPs use directly-connected topology, in which ISPs just connect to each other by direct link, not through a network switch?? If so, what's the percentage of this directly-connected case?
ISPs use a "direct link" (PNI) all the time to peer, and they don't need to be a member/customer of an internet exchange point to do so. In fact, the network you want to peer with might not be available at your local IXP even though they're in your local carrier hotel - then it becomes pretty much the only way to peer.
In London, the LINX offer switched *and* unswitched connectivity between members - you can rent fibers from them in order to perform PNI with other members. That the exchange offer this is unusual, and it's offered as an additional service, in order to smooth the process of organising interconnection between member organisations. [www.linx.net ]
LINX doesn't rent fibers. It's a one-time (NRC) fee for 8 pairs, which are patched to any other member on the service for free for life. (Although I don't know if they promise to keep it free forever, but it's been free for many years with no mention of it changing.)
We (LONAP) don't offer PNI services, and only offer the conventional switch ports, which members normally want so that they can get access to our peering lan and swap some traffic. [www.lonap.net] All exchanges offer this connectivity model. We offer private CUG and member-to-member private VLANs, which is similar, but still passes through the switch fabric.
I believe Exchange Point offers a PNI-like service over their network. But in general, an "Internet eXchange Provider" offers a shared switch. Anything else is really just a meet-me room. For instance, I wouldn't call Suite 1515 (formerly NYCC) an "IXP". -- TTFN, patrick
participants (7)
-
Andy Davidson
-
bmanning@vacation.karoshi.com
-
Greg VILLAIN
-
Kai Chen
-
Patrick W. Gilmore
-
Paul Vixie
-
Will Hargrave