On Wed, May 30, 2001 at 12:05:13AM -0400, Martin, Christian wrote:
> I am doing performance testing on a 7513 using VIP2-50s with PA-POS-OC3
> cards (8 all together). dCEF enabled, with 65000 flows running through it
> for a total of 600Mbps aggregate (70kpps with .1% loss). The CPU on the RSP
> is fine (0%). The VIPs are all at 99% doing RSB. The problem is this:
>
> 1) There isn't enough MEMD on RSP4 to switch packets agressively between
> VIPs, and
> 2) There isn't enough CyBUS bandwidth to do so either.
>
> What I am about to do is load up some RSP8s and rerun my test, as the 64MB
> of MEMD should help with the RSB drop issue. My question is this:
8MB of MEMD on RSP8, not 64MB. But it's still a lot more than 2MB.
>
> Does RSP8 add 'bandwidth' to the CyBUS or do we need VIP4-80s to get 4Gbps
> (I seem to remember something about a CzBUS)? I want to see how far the
> 7513 can go.
>
You need the VIP4s (50 or 80) to get more bandwidth.
> Anyway, the important point to remember is that true distributed switching
> on the RSP platforms only occurs between PA's in the same VIP (unlike the
> GSR). InterVIP switching requires the RSP to manage access to MEMD between
> VIPs (hence the high amount of RSB drops when the load is high).
>
Intra-vip switching isn't always purely local to the VIP. There are
some timing and starvation issues where with some PAs, we still send
traffic to MEMD. It's not often, and it only shows up in the
high-speed PAs, and I'm not sure if VIP4s alleviate that at all, but
that's how a few PAs on VIP2s behave.
eric
This archive was generated by hypermail 2b29 : Sun Aug 04 2002 - 04:12:39 EDT