[c-nsp] 3750X?

[Anton Kapela]

On Apr 15, 2010, at 3:58 AM, Gert Doering wrote:

> packets already under standard workload conditions).  Don't accept 
> handwaving and "QoS will fix this" arguments (as turning on QoS on the 
> 2960 will carve the tiny buffers into 4 times tiny/4 buffers, making the 
> problem *worse*...)

+1

If we looked only at the buffering configuration in the me3400, 2960, 3560, etc. it would seem that cisco is under the impression we're all doing single-MS-rtt's on our boxes, operating "metro area" networks in 2-mile square cities, and do not need (or even want the option to pay for) deeper queues for some ports. 

FWIW (and yea, I'm aware this is "C"-nsp), the "new" brocade/foundry CER2000 stuff is sporting 64/128/192 megabytes of shared packet buffer memory for the 24/28 + 2 10gig boxes, much like the 3550 was originally doing (i.e. shared buffer mem + per-port occupancy limits). Given that many of the "places" we tend to want a many-port gige or small 10 gig box, this could represent 100+ ms worth of buffering for a busy 'gige egress' when needed, handily permitting single-large tcp flows (or 10's of K's of multiplexed flows) to utilize all of the bandwidth/delay product. Perhaps ask them (your rep): "got something under $15K list that will come close to this?" 

Also, cer2k has 512k v4 tcam entries, dual ac/dc, consumes 1u of space, draws under 300 watts, and actually has working bgp/vpls/etc *today* -- not "somewhere over the me3400G rainbow."

-Tk