[j-nsp] Show Task Accounting

Pedro Roque Marques roque at juniper.net
Mon Aug 29 02:33:14 EDT 2005


India india wrote:
> Hi All,
>  
> In Juniper M40e... Is it possible to get the details on the maximum percentage of CPU Allocated for the 
> BGP Process.. I have configured 2400 E-BGP Sessions  in the Router and i can see always some sessions
> keep going down and come up automatically.
>  

There is no maximum of CPU for a particular process. As for why you are 
seing sessions drop, that doesn't automatically mean it is the BGP 
process not getting enough resources.

The best would be for you to contact j-tac about this so juniper could 
look at the details of your setup.

In case you don't wish to, this is what comes to the top of my head as 
possible diagnostics steps:

- sh system process extensive.
o is rpd stuck at 99% cpu ? Even if that is the case, that shouldn't 
necesarly mean that sessions would get dropped. "show task jobs" and 
"show task accounting" would give jtac and idea of what the box is doing.

- how many BGP groups do you have configured ? Most of the BGP 
processing scales in terms of the number of groups not sessions. 2400 
sessions should work 2400 groups will not.

- what sort of equipment do you have on the other side of the session ?
o When using testing equipment in labs it is quite often to see the 
router scale better than the testing equipment. Quite often the latter 
doesn't even have a real TCP stack.

- what error messsages do you get in the logs about session drops ? If 
you get holdtime expiry errors look at the TCP control block fields 
associated... e.g. sndcc == size of TCP send buffer. If this is not 
empty that usually means that rpd did send info to the TCP stack.

-  show pfe statistics traffic.
o Look under "local traffic statistics". An high number of input drops 
could cause TCP sessions to time out.

- show system statistics tcp (aka netstat)

Again, the best would be to contact j-tac... a box can get quite buzy w/ 
an high number of ebgp sessions. I also assume that this means 2400 
directly connected interfaces ? any link protocol running over those ? 
(e.g. ilmi).

btw: are you just trying to figure out what is the box's breaking point 
or do you have a real application in mind ?

regards,
   Pedro.


More information about the juniper-nsp mailing list