On Wed, Aug 25, 2021 at 02:01:26PM +0200, Kristjan Komlosi wrote: > On 24. 08. 21 21:59, Laura Smith wrote: > > Would be interesting to hear comments from the community on this comparison > > : https://elegantnetwork.github.io/posts/followup-measuring-BGP-stacks/ > > > > N.B. For the record, don't shoot the messenger, I had nothing to do with > > these tests, I just became aware of them via the BIRD list. I am > > particularly interested in the OpenBSD community comments given one person > > on the BIRD list had this to say of OpenBGPD: "OpenBGPd has always been a > > dog.". > > > > I'm no expert at all, but I'd imagine that OpenBGPD performs at least > somewhat differently on Linux, which seems to be what the author used in the > tests. My personal BGP server runs OpenBSD on a 512MB VPS, using about 150MB > of RAM with full IPv6 table and routing my traffic just fine, though I can > imagine the tables turning very quickly with lots of neighbors, as the > benchmark shows. I could try replicating their setup on an OpenBSD system, > but I don't have good enough hardware at hand at the moment.
The massive amount of memory used in OpenBGPD comes from the fact that unlike BIRD OpenBGPD runs with a full Adj-RIB-Out. The tests result in large amount of prefixes that need to be tracked. If you have 100 peers announcing 10000 random prefixes then you end up with 100 * 100 * 10000 = 100Million elements to manage. This is not a realistic test since in most cases the number of routes in the Adj-RIB-Out is limited (even on route servers). In the end for day to day use OpenBGPD performs well enough for many people. Future releases will focus more on performance and optimizing Adj-RIB-Out is on the list. -- :wq Claudio
