On 5/2/13 10:27 AM, Jeff Wheeler wrote:
On Wed, May 1, 2013 at 8:27 PM, ryanL <ryan.lan...@gmail.com> wrote:
i'm guessing this is a buffer thing, but i can't explain why it only
happens on my 1ge ports and not when i punt the traffic over an 10ge
Yes, it is a buffer thing. A 10GE interface is basically never going
to not have time to transmit frames unless it is receiving from 10 or
more 1GE interfaces at the same instant, steadily, for long enough to
fill the buffer; or there is at least one 10GE interface also talking
to it. On the other hand, two 1GE interfaces transmitting toward the
same out-going 1GE port can fill its buffer.
This is sometimes not obvious, because you look at the long-term
traffic and see a few hundred Mb/s, thinking, "why is there packet
loss?" You must keep in mind that the available buffer on modern ToR
switches is often less than 1ms worth of traffic.
ex4500 has like 4MB of shared buffer per PFE which is not really enough
to absorb a lot of microburst type activity.
The "buffer bloat" discussion of recent years has not done us any
favors. Many customers now think that buffers have historically been
too big. In fact, they were just often used incorrectly / configured
badly. Now we are not evaluating purchases based on having sufficient
buffer, so vendors have spent years developing products that ... lack
sufficient buffer.
Well it's a bit different circumstances between tiny CPE routers sitting
on dsl or cable endpoint and and a 48port 10Gbe switch. design-wise a
TOR has either whatever shared memory fits in the asic, or a large
external buffer. There's literally no options in between. so a 1/10Gb/s
TOR like the force10 s60 might have 2GB of shared packet buffer, while
an like an arista 7050s-64 would have 9MB for all the ports, assuming
you run it as all 10Gb/s rather than 100/1000/10000/40000 mixes of
ports it can cut-through-forward to every port which goes a long way
toward ameliorating your exposure to shallow buffers.
_______________________________________________
juniper-nsp mailing list juniper-nsp@puck.nether.net
https://puck.nether.net/mailman/listinfo/juniper-nsp
