[j-nsp] bgp metric-out igp and CPU utilization

[Daniel Verlouw]

Hi list,

before i open a tac case, wondering if anyone has seen something similar;

when we enable 'metric-out igp <offset> delay-med-update' towards a full-table downstream bgp customer (exporting ~400k prefixes), CPU % of rpd process spikes through the roof (mostly in kqread state), overall RE CPU % constantly stays well above 60% and BGP convergence seems to take forever in this state. And this is just on a single neighbor.
With 'metric-out <fixed value>' all is ok, and RE CPU % is in the 5-10% range.

'show task accounting' is showing high user run time for 'RT' and the BGP peer group on which we enabled the igp-based setting.
'rtsockmon -t rpd' appears to show no excessive route churn or oscillation for both igp-based and fixed value cases.


'show krt queue' is not showing any 'stuck' routes.
IGP (IS-IS) and RSVP-based LSPs are stable.


config bits;

[routing-options]
med-igp-update-interval 10;

[protocols bgp group IPTcust]
metric-out igp <offset> delay-med-update
or;
metric-out <fixed value>



This is on MX960 with 10.4R10.7. 

Any ideas?

--
Daniel.