As our border router we use a 1.7Ghz P4 machine with two upstream links, each with a BGP feed from the provider. For a year now that machine has been running pretty much without a problem, it's been running 2.8.10 for quite awhile now. Yesterday, after being up for about 2800 hours without a reboot, it began refusing to route to the outside world, it would only route via networks directly assigned to interfaces. Ok, we'd seen that before, it happens after the box has been up for a long time. Trying to access the routing->bgp menu hung, which is our standard test, the bgp process gets hung for some reason.

Since we had to reboot anyway, and 2.8.10 is getting rather old, we upgraded it to 2.8.26 before rebooting it. The box upgraded fine, rebooted, then things started getting strange.

The BGP feed from one provider downloaded in under two minutes, just as it usually does. The feed from the other provider though was pretty much useless. We'd get a few thousand routes in, the feed would hang and eventually disconnect, then connect and try to download the routes again. The entire time that 1.7G CPU was at 100% utilization. We rebooted a half dozen times, sometimes we'd get more routes before it hung, but never got the full tables from the second (slower) provider. We did notice that both sets of routes seemed to arrive pretty much normally until the first (faster) provider completed, then the second (slower) one hung and timed out every time.

Our next step was to downgrade the box, and we still had a copy of 2.8.17 in our file archives. After the downgrade both BGP feeds came up normally, though a little slow, it took 25 minutes for the slower feed to complete and once again the CPU was at 100% utilization. Once the initial downloads were complete, utilization dropped to the normal 5-10% for that box.

As best I recall from earlier discussions here in the forums, there aren't a lot of people running dual homed BGP, but for those who are, have you ever seen anything like this? What level O/S are you running?

It sure looks like something in the BGP code changed, and what ever the O/S is doing to merge those feeds from two different peers, it appears to be terribly inefficient. We're going to be hanging on to that copy of 2.8.17 because it sure appears dual homed BGP in later versions becomes useless.

Hello, follow this thread http://forum.mikrotik.com//viewtopic.ph ... t=bgp+keep
Something similar has happens to us.