Extreme BlackDiamond
Bradley Dunn
bradley at dunn.org
Mon Oct 13 22:49:42 UTC 2003
Steve Francis wrote:
>> BGP Scanner taking up close to 100% of CPU on a box periodically.
>> GSR doesn't seem to do it, but a buncha other cisco boxes do.
>> Its more irritating than anything else, especially when customers
>> complain
>> that when they traceroute they see ~200ms latency to the router...
>>
>>
> Doesn't happen here with MSFC2/SupII.
>
> Maybe just MSFC1's that are subject to that.
Every IOS-based device running BGP will have a BGP Scanner process that
wakes up once a minute and walks the BGP RIB checking that the next hops
are still valid. Whether it makes a noticeable impact on CPU utilization
depends on the platform and the size and distribution of the BGP table.
Obviously the more powerful the CPU the less the impact. In my
experience the distribution of the BGP table can also make a big
difference. Adding more specifics of a /8, /16, or /24 prefix seems to
have a disproportionate impact; my guess is it has something to do with
the data structure used to store the prefixes. (If they use a 256-way
mtrie like they do for CEF, more specifics of a /8, /16, or /24 would
require creation of an additional internal node.)
If you have a recent IOS that supports the 'show proc cpu history'
command, often the BGP Scanner spikes are quite obvious.
On platforms that do distributed forwarding, the spikes really only
affect traffic to/from the router, so additional latency will show up in
traceroutes or pings but forwarded traffic will not be affected. On
platforms that do centralized forwarding BGP Scanner can impact
forwarded traffic.
Bradley
More information about the NANOG
mailing list