[c-nsp] Weird Multicast microburst amplification issue

[Jeff Bacon]

This might sound incredibly silly, but in the case I've had better luck with 4900s. They use a single central buffer, so the buffer is effectively considerably deeper, depending on the specific workload of course. 

I've never been quite able to bring myself to use a dell blade chassis, and none of my people want to do HP. Plus we're playing with different 10G NICs for perf tuning. So, piles of 1Us and Aristas.

assuming you've got 6708/6716s (or are you using the 10Gs on the vs720?), you COULD attempt to use SRR QoS to attempt to knock off the edges of the peaks a little - say, cap at 500m-1Gbit for those flows to try and smooth out the microbursts a little on the 10g pipe. 

or, the gig ports on the supervisor IIRC have somewhat larger buffers. not sure how much help that is though, it's not like there's a lot of 'em. 

-bacon 


> -----Original Message-----
> From: Matthew Huff [mailto:mhuff at ox.com]
> Sent: Tuesday, December 13, 2011 9:49 AM
> To: Jeff Bacon; 'cisco-nsp at puck.nether.net'
> Subject: RE: Weird Multicast microburst amplification issue
> 
> I agree that the 10g->1G is probably the culprit amplifying the
> burst nature of the packets. At our colo at the NYSE in Mahwah we
> have implemented HP's flex fabric in blade chassis instead of doing
> what you are doing with the Arista's and it's working fine. We
> probably are going to have to do the same at our core datacenter.
> 
> http://h18004.www1.hp.com/products/quickspecs/13127_div/13127_div.h
> tml
> 
> Basically you connect N number of 10GB uplinks into the flex fabric
> and you can control how the bandwidth is allocated to each blade.
> 
> I was hoping to avoid having to do anything, but it looks like the
> data rates are killing the port buffers. I was hoping that the
> 6500/sup720 with 6748 would handle > 120Mbps, 12k pps multicast,
> but it doesn't look like it.