[rr-fs] current router scalability limits

[Howard C. Berkowitz]

At 6:05 PM -0800 3/25/04, Dima Krioukov wrote:
>i'm experiencing some difficulties trying to find
>a good set of references (preferably in conference
>proceedings/journals) for relatively accurate
>measurements of modern router scalability limits
>(the focus is on bgp).
>
>what's known to me so far is that:
>
>1) routers can hold for up to a million of fibs
>entries and do wire-speed lookups on them -- no problem.
>
>2) routers can hold for up to several millions
>of rib entries since they have a lot of memory --
>no problem.
>
>what's not known is even an estimate of how many rib
>entries they can hold under realistic load on both
>the control (bgp update traffic) and data (traffic
>to forward) planes.
>
>the latest results i could find so far are all not
>later than ~2001, all saying that the answer is not
>more than several hundred thousands before significant
>performance degradation.

Dima, I don't know if Alvaro Retana is on this list, so I don't want 
to speak for him if he is.  During our BMWG work, he contributed 
several scenarios about how the route mix in update had a significant 
impact on control plane load.  This was an interaction between the 
ordering (e.g., sorted by prefix length, linear expression of a tree 
of prefix and more-specifics) or randomness of the update stream, and 
the receiving data structure (hash table, linked list, etc,).  The 
number of NLRI per update also could be significant.

Others on the team experimented with their implementations and the 
interaction between different implementations. There were surprising 
results -- I seem to remember that one of the implementations 
converged faster, when receiving updates from a different 
implementation, than it did with another instance of the same code. 
Apparently the ordering of the other implementation's update was more 
optimal than the implementer's own ordering. I didn't do these tests 
myself, so I'd have to check which pairing it was. Various people 
tried different combinations of, at least, IOS, JunOS, Bay RS, 
NextHop GateD, and Zebra.

The methodology for exploring this is in our expired methodology 
draft, which we put on hold both because people went to other jobs or 
non-jobs, and we wanted final approval of our terminology draft.  The 
latter has been sitting in the IESG for some months, but I understand 
goes to IESG ballot on April 2.

Both the terminology and methodology grants are in the 
/internet-drafts directory on my anonymous ftp server at 
www.netcases.net.

>
>the central question seems to be how to introduce realistic
>load on the control plane of a dut, ie synthetic bgp traffic,
>but this all is work-in-progress (being done by olaf and anja,
>and few folks from bell-labs).
>
>is there anything i'm missing?
>
>thanks,
>--
>dima.
>http://www.caida.org/~dima/
>
>
>
>_______________________________________________
>rr-fs mailing list
>rr-fs at caida.org
>http://login.caida.org/mailman/listinfo/rr-fs