scharf at isc
Mar 21, 2012, 3:03 PM
Post #14 of 18
OK, I'll actually try to help a bit. My testing was for default free BGP
feeds (~360k routes with "typical" update rates) with full cross
distribution and route injection into linux (note that all these numbers
are for the control plane processing, traffic forwarding was not tested
at the same time for this work.) We found we ran out of CPU on the
unthreaded bgpd long before we ran out of memory. With our given
snapshot, we were able to get 18 peers with full cross feed (adding peer
18 resulted in >2 minutes of cpu time for bgpd before returning to a
more normal state.) RAM use was still under 4GB, but a 3GHz xeon was out
of gas. Take that as one data point. I know that there are quagga based
route servers with >100 peers out there, so that's another data point.
Ask the Cisco advocate to price out a system that has enough ram and cpu
to handle 15 default free peers with cross feed.
I have not done the detailed research for how memory access played into
it, but it seems like a threaded bgpd could greatly improve the numbers.
On 03/21/2012 10:27 AM, Steve Clark wrote:
> On 03/21/2012 01:21 PM, Tyler J. Wagner wrote:
>> On 2012-03-21 16:36, Steve Clark wrote:
>>> Is there a limit like 1, 2, 4 etc or not simple question?
>> I believe the limit is INT_MAX. RAM is the real limit.
> Hi Tyler,
> Thanks for your input.
> Stephen Clark
> Director of Technology
> Phone: 813-579-3200
> Fax: 813-882-0209
> Email: steve.clark [at] netwolves
> Quagga-users mailing list
> Quagga-users [at] lists