[c-nsp] Cisco ASR920-24SZ-IM BVI Feature Limitations

[James Jun]

On Sat, Jan 16, 2016 at 03:54:01PM +0200, Mark Tinka wrote:
>
> On 16/Jan/16 13:57, Eric Van Tol wrote:
> 
> >
> > We've been pretty happy with the ASR, especially the models with 4x10G on-board. The cost is significantly less than an ME3600, even with a full suite of licenses (Advanced IP Metro, all 10G ports, all GE ports), and the footprint is much smaller (well, more shallow). 
> 
> +1.
> 
> We've started rolling them out since last December, and so far so good.
>

+1 also, we have several ASR920-24SZ-IM's and 24SZ-M's out in the field and we're very happy with them.

Aside from LAG limitations (workaround solution was to not use them :-S), the only other issue I've run into is that default port buffer/queue sizes (48KB?) are rather small.  This is a slight annoyance since typical deployment of 920 has at least 2x 10GE feeding the 1GE revenue ports on the box.  As I understand, 920 only has 12MB shared buffer space so that probably explains it, but on default queue sizes, almost every 1GE end-user port (no traffic-shaping on user ports, just full-rate 1G port with 10G uplink) excessively collects output drops on practically most trivial IMIX usage.

For example, a FreeBSD box sitting with 1GE behind ASR920 just doing wget from a download mirror 50ms away records output drops on 920; whereas a 1GE port off of ASR9K or MX80 would not collect output drops for this type of usage.  Sure, it is reasonable to expect an end-user running Speedtest.net or watching Netflix spamming multiple flows to cause output drops, but not on single flow of download.

As a workaround, raising the queue-limit to 512 KB per 1G port dramatically gets rid of output drops for trivial traffic.  You should still see drops for longhaul bursty traffic overwhelming a 1GE interface when stepping down from 10G uplink, but that's pretty much a reasonable congestion at that point, so dropping packet is better.

512KB seems to be reasonable; 24x1GE * 512KB = 12.2MB, so we don't oversubscribe the global buffer space, and it's roughly ~4ms of output buffer per port.

!
class-map match-any cos_all
 match cos  0  1  2  3  4  5  6  7 
policy-map MC_1G_512kb
 class cos_all
  bandwidth percent 100 
  queue-limit 512000 bytes
!


James