Juniper BGP Convergence Time

Mailman List Mirror (Read Only)
1

Hello:

I'm running two Juniper MX104s. Each MX has 1 ISP connected running
BGP(full routes). iBGP is running between the routers via a two port 20G
lag. When one of the ISPs fails, it can take upwards of 2 minutes for
traffic to start flowing correctly. The router has the correct route in the
routing table, but it doesn't install it in the forwarding table for the
full two mins.

I have a few questions if anyone could answer them.

   - What would a usual convergence time be for this setup?
   - Is there anything I could do speed this process up? (I tried Multipath)
   - Any tips and tricks would be much appreciated

Thanks in Advance

2

The MX104 has a notoriously slow PPC-based RE unfortunately.

Josh

3

That's true. But are you using per-packet load balance policy in your configuration?

4

Hello:

I'm running two Juniper MX104s. Each MX has 1 ISP connected running
BGP(full routes). iBGP is running between the routers via a two port 20G
lag. When one of the ISPs fails, it can take upwards of 2 minutes for
traffic to start flowing correctly. The router has the correct route in the
routing table, but it doesn't install it in the forwarding table for the
full two mins.

I have a few questions if anyone could answer them.

   - What would a usual convergence time be for this setup?

With MX104, between 50 seconds and many minutes. The RE is not really
dimensioned for full tables, unfortunately.

   - Is there anything I could do speed this process up? (I tried Multipath)

Covering floating static route.

5

You sure it doesn't have something to do with 60 seconds * 3 = 180 secs of BGP neighbor Time out before it believes neighbor is dead and remove routes to that neighbor?

Aaron

6

Do you need full routes? What about just a default route from BGP?

Erich Kaiser
The Fusion Network
erich@gotfusion.net
Office: 815-570-3101

7

Have you checked your timeouts ?

-Ben

8

I could use static routes but I noticed since I moved to full routes I have
had a lot fewer customer complaints about latency(especially when it comes
to Voice and VPN traffic).

I wasn't using per-packet load balancing. I believe juniper default is per
IP.

My timers are as follows
Active Holdtime: 90
Keepalive Interval: 30

Would I be correct in thinking I need to contact my ISP to lower these
values?

An interesting note is when I had both ISPs connected into a single MX104
the failover was just a few seconds.

Thanks again.

9

A last resort route (default route) could still be good to take from your
ISP(s) even if you still do full routes, as the propagation is happening on
the internet side, you should at least have a path inbound through the
other provider. The default route at least would send the traffic out if
it does not see the route locally. Just an idea.

10

Erich,

Good Idea. I can't believe I didn't think of that earlier. Simple and
effective. I will go ahead and request the defaults from my ISP and update
the thread of the findings.

Thanks!

11

Ask if they will configure BFD for you. I’ve not found many transit providers that will, but it’s worth a shot and it will lower failure detection to circa 1 second.

12

I did exactly that when I called. One is looking into the other hasn't
called yet. I will let you know what they say.

13

The Juniper default is to not do ECMP at all. Only a single route is
programmed into the FIB for each prefix in your RIB. If you e.g. have
routes to 198.51.100.0/24 pointing to ten different ports, all traffic
to that entire /24 will go out over a single port, unless you have
explicitly enabled ECMP.

To enable ECMP, you need this:

    policy-options {
  policy-statement ecmp {
      then {
    load-balance per-packet;
      }
  }
    }
    routing-options {
  forwarding-table {
      export ecmp;
  }
    }

in your configuration. Note also that "per-packet" is a mis-nomer; it
is really "per flow", based on a hash of the L3/L4 headers.

'show route forwarding-table destination 198.51.100.0/24' shows if you
actually have multiple routes in your FIB.

  /Bellman

14

While we are on ECMP topic...

In L3VPN, when I've learned say, 3 different routes all using different MPLS tags to the 3 remote PE's, is there a way to ECMP hash across all of the paths to load balance?

Aaron

15

Just be aware of the impact a default route can have on your infrastructure, such as uRPF no longer works as expected as everything has a valid route.

16

Thomas,

Thanks for the info. This is probably why my multipath configuration wasn't
working as I thought it would. I will give this a test run also.

Mike,

Interesting thought. This would mean rpf-check wouldn't work on my outside
interfaces. Good to know.

17

Thomas,

Thanks for the info. This is probably why my multipath configuration wasn't
working as I thought it would. I will give this a test run also.

Mike,

Interesting thought. This would mean rpf-check wouldn't work on my outside
interfaces. Good to know.

Not necessarily that it doesn't work at all, but there are platform-specific differences in terms of loose vs. strict, whether the default route is considered in RPF evaluation, etc. From https://www.juniper.net/documentation/en_US/junos/topics/task/configuration/interfaces-configuring-unicast-rpf.html#jd0e50

18

You shouldn't need to contact your ISP on the lowered BGP timers as BGP
should establish based on the lowest value. That said, they may have a
value limit where anything lower than that, is set at your own risk.

You can look at running BFD over the BGP session as well. Technically it
has nothing to do with convergence, but it can quickly detect a down issue
and drop BGP right away.

19

We've tended to shy away from it, but we have 2 customers we've done it for.

Mark.

20

Hello again:

I've tried using the default route, adjusting bgp timers, and mutlipath.
Unfortunately, these changes haven't helped much. Juniper support hasn't
been very helpful also. Although, I think I might have found the solution.

https://www.juniper.net/documentation/en_US/junos/topics/topic-map/forwarding-indirect-next-hop.html

Let me know what you think.