[c-nsp] Input queue drops due to backplane congestion

[lee.e.rian at census.gov]

-----mack <mack at exchange.alphared.com> wrote: -----

>Date: 05/05/2008 04:03AM
>Subject: [c-nsp] Input queue drops due to backplane congestion
>
>We have a 6509 with sup720 3bxl running SXF11.
>We have a number of 6704-10GE line cards and
>A number of 6748-GE-TX line cards all with PFC3BXLs.
>
>When one (only one) of the 6704 fabric channels peaks at about
>80% utilization coming from the fabric we start to
>See packet loss on the GigE ports.  Even if the traffic
>Is confined to one GigE blade.
>
>We have run a large number of troubleshooting exercises
>Including replacing Sup units and line cards.  The problem
>Is definitely fabric congestion.  Our current solution is
>Simply to add another line card and move one of the lines
>On that fabric channel.
>
>This is somewhat different from the expected behavior
>Which would be drops occurring on the fabric channel counters.
>
>Has anyone else run into this problem?

Maybe.  Another person in the office has a TAC case open for input queue
drops on a 6500/sup720 and the TAC engineer came back with this:

 "With the WS-X6704-10GE blades, ports 1 & 2 belong to one
  fabric channel and 3 & 4 to another.  When traffic is
  switched between two ports leading to the same asic, you are
  limited to about 8 Gig per port.  If you are pushing this
  limit, it is possible the flow-control mechanism could induce
  packet drops. This is a hardware limitation."

Regards,
Lee