Re: [RAM] Number of DFZ routers

On May 28, 2007, at 9:10 AM, Ted Seely wrote:

> Curious as to where you may be taking this?

Working backwards, these are exactly the questions that you have to  
ask if you're trying to determine how large a lookup result needs to  
be.  See Robin's previous work on SRAM lookups.

On May 27, 2007, at 3:35 PM, Robin Whittle wrote:
> Can anyone suggest the range of number of interfaces these
> routers have?  For instance, what is the largest number of
> interfaces on an operational router?

Chris and others have pointed out a limitation in IOS.  Note that  
this is not shared by XR and presumably JunOS.  In particular, from a  
router architecture viewpoint, what's even more relevant is the  
potential number of interfaces that COULD be required from a system.

Now, the currently largest system that I know of that's being  
publicly discussed is the CRS-1.  That's 97Tbps, although other,  
larger systems have been discussed and will almost certainly be  
necessary in the distant future.  Not everyone agrees with this.

> What would be a typical range of interface numbers for
> transit routers?

Today, I'd say that interfaces in the DFZ range from DS-3 to OC-768.   
For the architectural bound, we can look at the CRS with T-3's,  
giving us around 97Tbps/45Mbps ~= 2e6 interfaces.

It should be noted that some interface types (Ethernet) are multi- 
access and require a router to store not only an interface, but next  
hop information.  A reasonable guess right now is to provision for  
about 35e3 next hops per Ethernet interface.  This allows the same  
hardware to be sold into most enterprise accounts.

Net that leaves you at around 7e10 next hops.  Based on these types  
of numbers, many folks have pointed out that distributed systems may  
be more scalable if they perform multiple lookup stages.  Not  
everyone agrees with this.  The cost of going to multiple lookups is  
that it requires additional high-speed lookup hardware, right at the  
point of the highest bandwidth in the architecture, the interface  
between the internal fabric and a chassis.

> In addition to direct forwarding of Internet traffic, and
> handling the communications paths between the routers
> themselves, do transit routers often perform other functions,
> such as handling MPLS?

As noted, yes, this is required.  Also, a slew of other features that  
have forwarding time impact are also practical requirements in this  
market space.  For example, ACLs, policy based forwarding,  
accounting, and QoS all have non-trivial impacts on the forwarding  
plane and may directly impact your lookup.  There are a number of  
considerations that must be taken into account.

Tony

>

_______________________________________________
RAM mailing list
RAM@iab.org
https://www1.ietf.org/mailman/listinfo/ram