BGP RIB Collection

Nick Hilliard nick at foobar.org
Wed Feb 27 12:42:01 UTC 2013


On 27/02/2013 03:51, Randy wrote:
> *received-routes*?
> If you still enable soft-reconfig-inbound on your routers(customer-facing sessions not withstanding), you most certainly hate your routers more than OP...;-)

it impacts memory, but if your management plane has enough memory to handle
it, it's a useful debugging tool.  For sure, it's the first thing I throw
out if the management plane RAM runs short.

SNMP polling of large router lists can work out as O(n^2) CPU usage if the
router stores the polled objects as linked lists or in some cases, in tree
structures.  This is because snmpgetnext cannot maintain a pointer to the
next object, which in some situations will mean a complete tree walk
operation.  So your CPU requirements will scale according to (size of
structure) * (average number of complete walks through the structure).  If
you're using linked lists, or have a naive tree implementation, "average
number of complete walks through the structure" = "size of structure" / 2
for a full tree walk.  I.e. you can require (n^2)/2 complete runs through
the structure in order to run a full snmp dump.  Obviously this isn't
always the case, but there are some well known examples of where it happens.

For all its faults, soft-reconfig-inbound only adds O(N) to RAM
requirements and almost nothing to CPU.

Nick

> ./Randy
> 
> --- On Tue, 2/26/13, Nick Hilliard <nick at foobar.org> wrote:
> 
>> From: Nick Hilliard <nick at foobar.org>
>> Subject: Re: BGP RIB Collection
>> To: "chip" <chip.gwyn at gmail.com>
>> Cc: "North American Network Operators Group" <nanog at nanog.org>
>> Date: Tuesday, February 26, 2013, 11:21 AM
>> On 26/02/2013 17:24, chip wrote:
>>> Currently I'm gathering this data via SNMP.
>>
>> whoa, you must really hate your router to do that to it.
>>
>>> While this works it has its draw backs, it
>>> takes approximately 20 minutes per view, its nowhere
>> near real-time, and
>>> I'm unable to gather information for IPv6.  SNMP,
>> however, is faster than
>>> screen scraping.  All of the XML based access
>> methods seem to take about
>>> the same time as well.
>>
>> cisco:
>> --
>> term len 0
>> show bgp ipv4 unicast neigh x.y.z.w received-routes
>> --
>>
>> juniper:
>> --
>> show route receive-protocol bgp x.y.z.w | no-more
>> --
>>
>> Easily scriptable using rancid or something similar. 
>> Of course, this sucks
>> because you're only seeing the route summary, not any of the
>> attributes.
>>
>>> project is still in its infancy.  BMP seems to be
>> a good solution but I've
>>> not found a working client implementation yet.  I
>> see that you can actually
>>> configure this on some Juniper gear but I can't seem to
>> locate a client to
>>> ingest the data the router produces.
>>
>> Can you provide a list of the clients that you have
>> tried?  It would save
>> people the effort of going through them and finding out the
>> same things as
>> you did.
>>
>> Nick
>>
>>
>>
>>
> 





More information about the NANOG mailing list