If you have full use of the fiber across which you move your packets you can create a mesh of PVCs directly connecting each router thus decreasing the hop count. The same can be done when using an existing cell-relay cloud, but you pay on a per pvc basis so the benefit needs to be weighed against the cost. Of course the packets still flow along the same physical path and in a wide area network the time in transit will be more significant than the time to get through the routers.
Jim
Nope, remember - there is no magic. Any mesh of PVCs that one makes over a switched network must reflect the toplogy of that network, and one can set up a matching set of active routing sessions and route weights which will cause traffic to flow the same way. Yes, the switches are a bit faster and have less to do. Data moves through them in a few ms less per point. But as you said (and as I said in our discussion in NYC), relative to any distance, the speed of light guarantees that you won't notice the difference. The question is: Will there be routers available that can make IP routing decisions based on 40-60kroutes and move 2-3 OC3s worth of bidirectional traffic? The building of the configs to have a routed network work the same as a switched ATM one can be automated, but it's true that it *can be* easier to see what's going on in a large-scale switched network. Avi
Nope, remember - there is no magic. Any mesh of PVCs that one makes over a switched network must reflect the toplogy of that network, and one can set up a matching set of active routing sessions and route weights which will cause traffic to flow the same way.
Yes, the switches are a bit faster and have less to do. Data moves through them in a few ms less per point. But as you said (and as I said in our discussion in NYC), relative to any distance, the speed of light guarantees that you won't notice the difference.
The question is: Will there be routers available that can make IP routing decisions based on 40-60kroutes and move 2-3 OC3s worth of bidirectional traffic? The building of the configs to have a routed network work the same as a switched ATM one can be automated, but it's true that it *can be* easier to see what's going on in a large-scale switched network.
Avi
Uh, yes. Like today. See NetStar (recently acquired by ASCEND). Their hardware does nothing *slower* than DS-3 right now, and updates the route forwarding table 20x/second. :-) This thing is a monster, and will make the so-called "state of the art" 7513s look like dogmeat when in full production (presuming the BGP issues get completely resolved). Coming soon to a network near you. -- -- Karl Denninger (karl@MCS.Net)| MCSNet - The Finest Internet Connectivity http://www.mcs.net/~karl | T1 from $600 monthly; speeds to DS-3 available | 23 Chicagoland Prefixes, 13 ISDN, much more Voice: [+1 312 803-MCS1] | Email to "info@mcs.net" WWW: http://www.mcs.net/ Fax: [+1 312 248-9865] | Home of Chicago's only FULL Clarinet feed!
Uh, yes. Like today.
See NetStar (recently acquired by ASCEND).
Their hardware does nothing *slower* than DS-3 right now, and updates the route forwarding table 20x/second. :-)
This thing is a monster, and will make the so-called "state of the art" 7513s look like dogmeat when in full production (presuming the BGP issues get completely resolved).
Coming soon to a network near you.
Karl Denninger (karl@MCS.Net)| MCSNet - The Finest Internet Connectivity
What's the matter, Karl, don't like gated for running BGP? Avi
Uh, yes. Like today.
See NetStar (recently acquired by ASCEND).
Their hardware does nothing *slower* than DS-3 right now, and updates the route forwarding table 20x/second. :-)
This thing is a monster, and will make the so-called "state of the art" 7513s look like dogmeat when in full production (presuming the BGP issues get completely resolved).
Coming soon to a network near you.
Karl Denninger (karl@MCS.Net)| MCSNet - The Finest Internet Connectivity
What's the matter, Karl, don't like gated for running BGP?
Avi
Not fast enough when you're talking about OC3s ;-). -- -- Karl Denninger (karl@MCS.Net)| MCSNet - The Finest Internet Connectivity http://www.mcs.net/~karl | T1 from $600 monthly; speeds to DS-3 available | 23 Chicagoland Prefixes, 13 ISDN, much more Voice: [+1 312 803-MCS1] | Email to "info@mcs.net" WWW: http://www.mcs.net/ Fax: [+1 312 248-9865] | Home of Chicago's only FULL Clarinet feed!
Not fast enough when you're talking about OC3s ;-).
Karl Denninger (karl@MCS.Net)| MCSNet - The Finest Internet Connectivity
Gated doesn't have to be faster when the router is routing OC3s than when it's routing 56ks... I assume that the mondo-routers have a switching engine, and the only job of the main CPU/gated combo is to feed the current idea of the best next-hop/interface to the engine. Avi
Not fast enough when you're talking about OC3s ;-).
Karl Denninger (karl@MCS.Net)| MCSNet - The Finest Internet Connectivity
Gated doesn't have to be faster when the router is routing OC3s than when it's routing 56ks...
I assume that the mondo-routers have a switching engine, and the only job of the main CPU/gated combo is to feed the current idea of the best next-hop/interface to the engine.
Avi
This is true. But show me a model in which this works and which can update that forwarding table 20x/second (at full load WHILE switching packets). That's what Netstar claims. CISCO can't even *dream* of getting close to this level of performance. -- -- Karl Denninger (karl@MCS.Net)| MCSNet - The Finest Internet Connectivity http://www.mcs.net/~karl | T1 from $600 monthly; speeds to DS-3 available | 23 Chicagoland Prefixes, 13 ISDN, much more Voice: [+1 312 803-MCS1] | Email to "info@mcs.net" WWW: http://www.mcs.net/ Fax: [+1 312 248-9865] | Home of Chicago's only FULL Clarinet feed!
This is true. But show me a model in which this works and which can update that forwarding table 20x/second (at full load WHILE switching packets).
That's what Netstar claims.
CISCO can't even *dream* of getting close to this level of performance.
Karl Denninger (karl@MCS.Net)| MCSNet - The Finest Internet Connectivity
I think I'd be happy if it could update the entire forwarding table (only) a few times a second, but could switch a few gigabits worth of IP packets per second. Avi
This is true. But show me a model in which this works and which can update that forwarding table 20x/second (at full load WHILE switching packets).
That's what Netstar claims.
CISCO can't even *dream* of getting close to this level of performance.
Karl Denninger (karl@MCS.Net)| MCSNet - The Finest Internet Connectivity
I think I'd be happy if it could update the entire forwarding table (only) a few times a second, but could switch a few gigabits worth of IP packets per second.
Avi
Uh, the Netstar box *IS* called the "Gigarouter". There's a reason for that ;-) - -- Karl Denninger (karl@MCS.Net)| MCSNet - The Finest Internet Connectivity http://www.mcs.net/~karl | T1 from $600 monthly; speeds to DS-3 available | 23 Chicagoland Prefixes, 13 ISDN, much more Voice: [+1 312 803-MCS1] | Email to "info@mcs.net" WWW: http://www.mcs.net/ Fax: [+1 312 248-9865] | Home of Chicago's only FULL Clarinet feed!
This is true. But show me a model in which this works and which can update that forwarding table 20x/second (at full load WHILE switching packets).
That's what Netstar claims.
CISCO can't even *dream* of getting close to this level of performance.
They don't have to. They're selling routers. Maybe when Netstar's can be deployed, there'll be something to shout about. Until then, it just plain isn't worth arguing about. Dave -- Dave Siegel Sr. Network Engineer, RTD Systems & Networking (520)623-9663 x130 Network Consultant -- Regional/National NSPs dsiegel@rtd.com User Tracking & Acctg -- "Written by an ISP, http://www.rtd.com/~dsiegel/ for an ISP."
On Fri, 19 Jul 1996, Avi Freedman wrote:
I assume that the mondo-routers have a switching engine, and the only job of the main CPU/gated combo is to feed the current idea of the best next-hop/interface to the engine.
Current mondo-routers like Netstar uses a switch fabric to parallelise switching. It becomes pretty obvious that at OC-N speeds current serial backplane architecture doesn't scale Handling routing updates, unlike switching packets, is more a function of CPU and software routines. -dorian
In message <m0uhNhQ-000IDOC@venus.mcs.com>, "Karl Denninger, MCSNet" writes:
What's the matter, Karl, don't like gated for running BGP?
Avi
Not fast enough when you're talking about OC3s ;-).
Karl, Since when does a BGP implementation have to go any faster to support higher bandwidth IP forwarding. The amount of route change isn't going to go up any if you add an OC48 to your network. Curtis
Karl, We have been running NetStars in the vBNS for ~1 year. While the h/w is impressive, the GigaRouter s/w still aspires to the level of Cisco's. And no, I don't mean gated. As to your comment "(presuming the BGP issues get completely resolved)."--we have yet to take full routing in our production routers, but we are close. We have had some problems with the RMS (486 running BSDI & gated) being able to distrubute the forwarding table to the GigaRouter linecards (FDDI, ATM/OC-3c, & HiPPI) during start-up. New s/w addresses this pacing issue, and we expect to deploy BGP on our GigaRouters very soon. OSPF works very well on these boxes, and considering that when the NetStars were first deployed everything was static, we have made some progress :) Also, we look forward to getting our hands on some ATM/OC-12c cards this fall. The NetStar is an interesting router that has made a lot of progress, and I hope they stay in the race. Regards, Randy On Fri, 19 Jul 1996, Karl Denninger, MCSNet wrote:
Nope, remember - there is no magic. Any mesh of PVCs that one makes over a switched network must reflect the toplogy of that network, and one can set up a matching set of active routing sessions and route weights which will cause traffic to flow the same way.
Yes, the switches are a bit faster and have less to do. Data moves through them in a few ms less per point. But as you said (and as I said in our discussion in NYC), relative to any distance, the speed of light guarantees that you won't notice the difference.
The question is: Will there be routers available that can make IP routing decisions based on 40-60kroutes and move 2-3 OC3s worth of bidirectional traffic? The building of the configs to have a routed network work the same as a switched ATM one can be automated, but it's true that it *can be* easier to see what's going on in a large-scale switched network.
Avi
Uh, yes. Like today.
See NetStar (recently acquired by ASCEND).
Their hardware does nothing *slower* than DS-3 right now, and updates the route forwarding table 20x/second. :-)
This thing is a monster, and will make the so-called "state of the art" 7513s look like dogmeat when in full production (presuming the BGP issues get completely resolved).
Coming soon to a network near you.
-- -- Karl Denninger (karl@MCS.Net)| MCSNet - The Finest Internet Connectivity http://www.mcs.net/~karl | T1 from $600 monthly; speeds to DS-3 available | 23 Chicagoland Prefixes, 13 ISDN, much more Voice: [+1 312 803-MCS1] | Email to "info@mcs.net" WWW: http://www.mcs.net/ Fax: [+1 312 248-9865] | Home of Chicago's only FULL Clarinet feed!
------------------------------------------------------------------------ Randolph C. Nicklas MCI Telecommunications vBNS Engineering 2100 Reston Parkway Phone: 703.715.7099 Reston, VA 22091 Fax: 703.715.7066 Pager: 1-800-SKY-8888 PIN 902418
Here is the senario I was thinking about Case 1 R1 --- R2 --- R3 Case 2 R1 --- SW --- R3 | | R2 Assume that the wide area segments are the same in both cases and that SW and R2 are collocated. In case 1 there are 2 router hops between R1 and R3. In case 2, given a full mesh of PVCs, R1 and R2 are only one hop apart. This does not imply that the traffic flows any differently relative to the physical paths taken, but if I do a traceroute I think it will look different. Am I missing the point here? Jim
If you have full use of the fiber across which you move your packets you can create a mesh of PVCs directly connecting each router thus decreasing the hop count. The same can be done when using an existing cell-relay cloud, but you pay on a per pvc basis so the benefit needs to be weighed against the cost. Of course the packets still flow along the same physical path and in a wide area network the time in transit will be more significant than the time to get through the routers.
Jim
Nope, remember - there is no magic. Any mesh of PVCs that one makes over a switched network must reflect the toplogy of that network, and one can set up a matching set of active routing sessions and route weights which will cause traffic to flow the same way.
Yes, the switches are a bit faster and have less to do. Data moves through them in a few ms less per point. But as you said (and as I said in our discussion in NYC), relative to any distance, the speed of light guarantees that you won't notice the difference.
The question is: Will there be routers available that can make IP routing decisions based on 40-60kroutes and move 2-3 OC3s worth of bidirectional traffic? The building of the configs to have a routed network work the same as a switched ATM one can be automated, but it's true that it *can be* easier to see what's going on in a large-scale switched network.
Avi
Here is the senario I was thinking about
Case 1
R1 --- R2 --- R3
Case 2
R1 --- SW --- R3 | | R2
Assume that the wide area segments are the same in both cases and that SW and R2 are collocated. In case 1 there are 2 router hops between R1 and R3. In case 2, given a full mesh of PVCs, R1 and R2 are only one hop apart. This does not imply that the traffic flows any differently relative to the physical paths taken, but if I do a traceroute I think it will look different. Am I missing the point here?
Jim
No, you're not missing the point. My point is just that you can do the same things as well with routers, assuming the routers can push the amount of traffic you need to push. (And I think a 7507 or so can push 2-3 OC3s, though I could be wrong). My point is just basically a "so what if the IP-level path is shorter"? I fail to see why that's a Good Thing, especially since it hides information that I want to see... Acceptable arguments are: o Switches can handle more throughput o Switched networks are easier for humans (or at least, humans without huge brians && many internal registers) to design/debug/maintain. o Switches are requied because without going to a smaller decision space (re: how much work per unit the switch/router has to do to decide where to forward the data unit off to), existing routed technology can't support moving IP packets at the desired speed while having to evaluate the 40-60k routes we'll have soon. But not "my traceroutes are shorter". Anyway, you (Jim) run a very good, reliable network using ATM technology. I suppose I shouldn't prickle quite so easily, but I feel there are a lot of misconceptions out there... Avi
Here is the senario I was thinking about
Case 1
R1 --- R2 --- R3
Case 2
R1 --- SW --- R3 | | R2
Assume that the wide area segments are the same in both cases and that SW and R2 are collocated. In case 1 there are 2 router hops between R1 and R3. In case 2, given a full mesh of PVCs, R1 and R2 are only one hop apart. This does not imply that the traffic flows any differently relative to the physical paths taken, but if I do a traceroute I think it will look different. Am I missing the point here?
Jim
No, you're not missing the point. My point is just that you can do the same things as well with routers, assuming the routers can push the amount of traffic you need to push. (And I think a 7507 or so can push 2-3 OC3s, though I could be wrong).
My point is just basically a "so what if the IP-level path is shorter"? I fail to see why that's a Good Thing, especially since it hides information that I want to see...
Acceptable arguments are: o Switches can handle more throughput o Switched networks are easier for humans (or at least, humans without huge brians && many internal registers) to design/debug/maintain. o Switches are requied because without going to a smaller decision space (re: how much work per unit the switch/router has to do to decide where to forward the data unit off to), existing routed technology can't support moving IP packets at the desired speed while having to evaluate the 40-60k routes we'll have soon.
But not "my traceroutes are shorter".
Anyway, you (Jim) run a very good, reliable network using ATM technology. I suppose I shouldn't prickle quite so easily, but I feel there are a lot of misconceptions out there...
Avi
I don't disagree with anything you have said. I don't think "too many hops" is an interesting criticism of an NSPs network. All I was saying is that it is possible, using switches, to reduce your hop count. I am not arguing the merits of doing so. Today, regardless of how fast switches can push cells, if your network is IP based you need to traverse routers. When the routers get overloaded you must either install more routers or faster routers. Jim Ps. I don't take it personally Avi. I only argue with you because when I do I often learn something.
I don't disagree with anything you have said. I don't think "too many hops" is an interesting criticism of an NSPs network. All I was saying is that it is possible, using switches, to reduce your hop count. I am not arguing the merits of doing so. Today, regardless of how fast switches can push cells, if your network is IP based you need to traverse routers. When the routers get overloaded you must either install more routers or faster routers.
Hmm, I would say: Today, regardless of whether you use switches or routers or both, if you're not making effective/correct internal-WAN load-balancing decisions, you're screwed (assuming you have enough internal redundancy/ bandwidth to run your network even with 100% optimal routing decisions).
Jim
Ps. I don't take it personally Avi. I only argue with you because when I do I often learn something.
Good, I was going to send you private e-mail to make clear that I was just on a religious ATM vs. Routed network thing - nothing personal. I'll still come down to visit after ISPCON. And I incuded nanog@ to get the opinion of others... Avi
I don't disagree with anything you have said. I don't think "too many hops" is an interesting criticism of an NSPs network. All I was saying is that it is possible, using switches, to reduce your hop count. I am not arguing the merits of doing so. Today, regardless of how fast switches can push cells, if your network is IP based you need to traverse routers. When the routers get overloaded you must either install more routers or faster routers.
an average number of hops is no big deal...too many hops, and the ttls get too high, so there are certainly instances where some level of meshing is necessary (but meshing doesn't imply switching with ATM or Frame..) Dave -- Dave Siegel Sr. Network Engineer, RTD Systems & Networking (520)623-9663 x130 Network Consultant -- Regional/National NSPs dsiegel@rtd.com User Tracking & Acctg -- "Written by an ISP, http://www.rtd.com/~dsiegel/ for an ISP."
Acceptable arguments are:
o Switches can handle more throughput
That's difficult to quantify in theory *or* practice.
o Switched networks are easier for humans (or at least, humans without huge brians && many internal registers) to design/debug/maintain.
More levels of indirection does not mean it's easier for humans necessarily. In fact, there are many more nobs to miss, and more places for error to be introduced into your engineering model.
o Switches are requied because without going to a smaller decision space (re: how much work per unit the switch/router has to do to decide where to forward the data unit off to), existing routed technology can't support moving IP packets at the desired speed while having to evaluate the 40-60k routes we'll have soon.
But stuff that's in development sure can.. Dave -- Dave Siegel Sr. Network Engineer, RTD Systems & Networking (520)623-9663 x130 Network Consultant -- Regional/National NSPs dsiegel@rtd.com User Tracking & Acctg -- "Written by an ISP, http://www.rtd.com/~dsiegel/ for an ISP."
(Agislist taken off cc: list)
Acceptable arguments are:
o Switches can handle more throughput
That's difficult to quantify in theory *or* practice.
The practice is that there are switches that will do multiple full OC3s and few routers that can, no?
Empty cardboard boxes "will do" multiple full OC3c's too. How are you measuring quality? How do the kinds services your clients want affect how that performance is rated? Dave -- Dave Siegel Sr. Network Engineer, RTD Systems & Networking (520)623-9663 x130 Network Consultant -- Regional/National NSPs dsiegel@rtd.com User Tracking & Acctg -- "Written by an ISP, http://www.rtd.com/~dsiegel/ for an ISP."
The question is: Will there be routers available that can make IP routing decisions based on 40-60kroutes and move 2-3 OC3s worth of bidirectional traffic?
There are now. See www.netstar.com for details. Last time I looked their route table could grow to a couple hundred K's of routes, too. Back when I used one the slowest media card you could purchase was OC-3... ...arun --------------------------------------------------------------------------- Arun Welch 5000 Arlington Centre Blvd Lead Software Engineer Columbus, OH 43220 CompuServe awelch@csi.compuserve.com
participants (9)
-
Arun Welch -
Avi Freedman -
Curtis Villamizar -
Dave Siegel -
Dorian R. Kim -
Jim Van Baalen -
Karl Denninger, MCSNet -
Perry E. Metzger -
Randolph C. Nicklas