[j-nsp] router dropping packets where there is no reason anywhereto be seen

[Rafal Szarecki (WA/EPO)]

Alexander,

I not katch what exactly problem you have but please see notes below:

The policing are executed on Internet Processor. All below M40e has only one IP II, so policing of ae interface is acurate.

On M160 traffib from single ae interface is spread in per packet roud robin to 4 SFM (each host single IP II). In fact if you configure 10Mbps policer, on each SFP 2,5Mbps policer is implemented.

On M320 and T-series Internet Processor  are located on PFE complex on FPC. Single for FPC except T640 FPC type 3 where two IP exist serving PIC0-1 and PIC2-3 respectivly. If ae member interfaces are served by different  PFE complexes, then again on each PFE policing rate is set proportionaly to number of members of ae group served by this PFE.
If traffic will be distributed equaly over ae members, then policing will be acurate. But this is not always true. The recommendation is to run per-flow load balancing which as deep hash as possible - this should help with egress policing. To the same on BD to improove ingress policing.

I do not know case when defaulr QoS settings introduce packetdrops. Of course if no congestion. By default all packets goes to queue 0 (Best-effort) regardless of DSCP/ToS/.1p/EXP. Only one exception is IPprec 7 and 6, which are placed to queue3 (5% rate, 5% buffer space -->200ms) , because this values are Network controll - routing.
QoS for ae was very week till latest JUNOS (7.3 or 7.4 - I'm not sure)

Rafał Szarecki JNCIE

skype me <callto://Rafal_Szarecki/> 


 

> -----Original Message-----
> From: juniper-nsp-bounces at puck.nether.net 
> [mailto:juniper-nsp-bounces at puck.nether.net] On Behalf Of 
> Alexander Koch
> Sent: Thursday, November 03, 2005 9:15 AM
> To: juniper-nsp at puck.nether.net
> Subject: [j-nsp] router dropping packets where there is no 
> reason anywhereto be seen
> 
> Folks,
> 
> when asked to debug a problem for a close friend I stumbled 
> over what they have. Consider a M160 router, an OC48, some ae 
> to some Black Diamond switches, and some traffic, but all 
> around 300-400 Mbit/s, and one GE at 700.
> 
> I was wondering as I had some nasty things popping back up, 
> namely that due to the different architecture policers on a 
> M20/M40 (which do what you tell them) where quite hazardous 
> on our M320s.
> 
> Also I seem to remember (correctly, I hope) that when no 
> class-of-service is configured some default stuff kicks in, 
> and there is an occasional drop here and there, just because 
> of the default stuff.
> 
> The symptoms are easy, packetloss on the box and especially 
> so through the box. Debugging this can be nasty, yes. What 
> caveats are there for the M160 and default QoS and default 
> policers for the RE (lo0). Also what is the default ARP 
> policer, as that might have been an issue in the past as 
> well. 20K MACs are now handled through the BD switches, mind you.
> 
> Does anything I write here say anything to anyone? Would QoS 
> help? I would say yes here, but our backbone ourselves is so
> feature- rich that I might be comparing Apples to Dells. ;->
> 
> Thanks for any feedback.
> 
> Regards,
> Alexander
> 
> _______________________________________________
> juniper-nsp mailing list juniper-nsp at puck.nether.net 
> http://puck.nether.net/mailman/listinfo/juniper-nsp
>