[c-nsp] 4x10G Etherchannel overruns

[James Bensley]

On 3 March 2017 at 17:04, Peter Kranz <pkranz at unwiredltd.com> wrote:
> On a WS-X6908-10G DCEF2T line card with SUP2T's, I ran into overruns
> yesterday on a 4x10G etherchannel that I am at a loss to resolve:
>
>
>
> Constantly increasing overrun counter:
>
>    6418130558941 packets input, 9277559958229871 bytes, 0 no buffer
>
>      Received 668274 broadcasts (0 IP multicasts)
>
>      0 runts, 190 giants, 0 throttles
>
>      192 input errors, 1 CRC, 0 frame, 51591389 overrun, 0 ignored
>
>
>
> Latency into the router rose by 40ms when these overrun's started to appear
>
>
>
> This happened at a BW of ~28 Gbps
>
>
>
> I've built the etherchannel in this manner:
>
>
>
> Index   Load      Port          EC state       No of bits
>
> ------+------+------------+------------------+-----------
>
> 0      0A            Te1/1             Active   2
>
> 3      81            Te1/2             Active   2
>
> 1      60            Te1/3             Active   2
>
> 2      14            Te1/4             Active   2
>
>
>
> Is it necessary to instead stagger 1/1, 1/3, 1/5, 1/7 to spread the load
> across the card ASICs? I didn't think the WS-X6908 was an oversubscribed
> card so didn't bother initially.


This is supposed to be a non-blocking card (no oversubscription, 1
port per ASIC, line rate 80Gbps at 128 byte packets):

http://www.cisco.com/c/en/us/products/collateral/switches/catalyst-6500-series-switches/white_paper_c11-652042.html

Just double checking on one here, this is a WS-X6908-10G DCEF-2T in a
6509-E with SUP-2T, there are 8x 10G R2D2 ASICs (one per port), 4x
Metro Argos etc, so I think there should be no oversubscription here:

6513E#show asic-version switch 1 slot 1
Module in switch 1, slot 1 has 7 type(s) of ASICs
ASIC Name      Count      Version
METRO_ARGOS          4      (4.0)
METRO_KRYPTON          4      (4.0)
MALFRA          2      (16.4)
SSA          4      (9.0)
SANTA_MONICA          1      (1.0)
GANITA          8      (2.0)
R2D2          8      (5.0)


Can you provide the same "show int" for each member link as well as
the bundle? Do you see overruns on member links or the port-channel
interface only?


Cheers,
James.