[c-nsp] 2960 -> 4948 - no more drops :)

[Phil Mayers]

On 19/02/13 11:29, Reuben Farrelly wrote:
> On 19/02/2013 9:21 PM, Peter Rathlev wrote:
>>> This is a classic example of when a Gig port in name is not a Gig port
>>> in throughput, ie it may link up at that speed but you'd be lucky to get
>>> the rated throughput in all but ideal circumstances.
>>
>> Funny thing is that many lower end switches (i.e. cheaper) have better
>> buffer characteristics than a 2960 or similar. And even though the 2960
>> is cheap compared to other high end switches it isn't compared to
>> gigabit switches in general.
>
> Notwithstanding that this is a Cisco list, is this limited buffer
> problem something common to many other vendor's lower end switches too?
>   Do switches from vendor J, Mikrotik and HP also have a similar problem?

Many do, because many are built on the same hardware these days. A lot 
of the 1U gig/10gig devices are all pretty much standard Broadcom designs.

The situation gets slightly more complex with some newer switches that 
are cut-through, where the buffers are only eaten into if there's 
contention or a port-speed mismatch between ingress and egress.

>
> It's obviously used as a differentiator of high and low end switches for
> Cisco, but I've often wondered if it's used as a deliberate "cripple" to
> the kit, surely it can't be THAT hard and expensive to engineer
> increased shared buffer space in a low end chassis from 12M to, say 60M?

Well, the RAM needs to be pretty fast, so it's not cheap, but the ASIC 
also needs to be able to *address* that size of RAM; it might have e.g. 
a 24-bit address space.

The buffer size thing is a pain, but equally over-buffering is also bad; 
there's no easy solution AFAICT.