[j-nsp] l2circuit communities

[Phil Bedard]

On May 25, 2010, at 3:21 AM, Richard A Steenbergen wrote:

> On Mon, May 24, 2010 at 07:46:17PM -0400, Phil Bedard wrote:
>> A little different scenario but I'm using CBF with a cos next-hop-map
>> to set specific lsp-next-hops for CoS classes, also using autobw, and
>> I'm not seeing similar behavior. 
>> 
>> One thing I noticed is while the router is doing the MBB/re-signal the
>> "ActiveRoute" value will drop to 0, but then it immediately goes back
>> up to the prior value, so it makes you wonder what's going on behind
>> the scenes. 
>> 
>> I'm using 9.3R3.8, hope thing isn't something introduced later... 
>> 
>> Are your paths actually changing output interfaces/paths or is it just
>> a BW re-signal? 
> 
> Just doing a bandwidth resignal, nothing else is changing. I see the 
> same behavior with the ActiveRoute count, it drops to 0 for about 14 
> seconds after the resignal, then pops back up to where it was before.
> 
> I don't have anything in the 9.3's to test, but 9.6S5 is definitely
> doing the same thing. I also confirmed that the route changes in
> rtsockmon are from routes pointing to LSPs which just resignaled, which
> stops as soon as the install-nexthop config goes away. I'm not
> particularly interested in going to a full cos next-hop-map setup, but
> this is definitely broken as-is. :)
> 

I ran some tests with autobw, cos next-hop map, and your method of installing lsp next-hops 
based on communities to see what the CPU looked like.  I tested this against 9.3R3.8 on an M320 and 10.2R1.3 on an MX960.  

With the cos next-hop map the RPD CPU usage was roughly double what it was without the cos next-hop map.  Without it, the M320 hit 
10-12% CPU during a re-signal and the MX960 was 15-18%.   With the cos next-hop-map I'd see it jump to ~24% on the M320 and hit a 
peak of 40% on the MX960.  
 
Both normal autobw and the cos next-hop-map consistently used more CPU on the MX960 with the RE-2000 than 
the M320 with the RE-1600.  One thing I noticed is on the 10.2 MX960 box I saw more change messages via rtsockmon 
than with the 9.3R3.8 box.  I had static routes pointed to the LSPs in order to vary the BW on them and every time 
one re-signaled I'd see a nh change for the static route.  I also saw indirect nh changes.   On the 9.3R3.8 box I 
never saw the change messages.  

The test installing the lsp next-hop via communities would bring both boxes up to 70-95% during LSP re-signals and 
result in seeing a change for each individual route using the LSP.   

The tests were done with 400k BGP prefixes reachable via a set of 10 LSPs with autobw enabled.  Traffic was varied 
across the LSPs to get them to re-signal every 5 minutes.   

Phil