[f-nsp] Features on Brocade Ethernet platforms

[George B.]

On Sun, Mar 13, 2011 at 12:35 AM, Robert Hass <robhass at gmail.com> wrote:

> On Sun, Mar 13, 2011 at 7:55 AM, George B. <georgeb at gmail.com> wrote:
>
> > An alternative to the TurboIron 24x might be something like the Arista
> 7100
> > series depending on what features you need.  They produce an ER optic.
>
> George, and how about preventing micro-bursts mentioned by Nick at
> Arista 7100 gear ? How deep buffers it has ? How about pricing
> comparing to TI24X (it's $12k GPL). I just need 8-16 10GE ports in
> 1-2U form factor.
>
> Robert
>

I would say first go here:


http://www.aristanetworks.com/en/products/7100series

Notice the PDF link to "Myths about microbursts"

Here is a link to the data sheet:

http://www.aristanetworks.com/media/system/pdf/Datasheets/7100_Datasheet.pdf

And if you have been following Jim Getty's stuff ... big buffers actually
probably hurt performance rather than help, particularly if there is
congestion in the path anywhere (including outside your network in the
Internet path).  The "every packet is sacred" approach to huge buffers to
prevent packet loss in the case of congestion probably hurts performance
more than it helps.  Dropping a packet and having a flow back off a little
bit is how TCP was designed to deal with congestion.  Having several seconds
worth of traffic sitting in various buffers along the path means that by the
time you learn that a path is congested, it might be too late to really do
anything about it.  So you end up losing a packet, you back off, but now you
are still losing packets that were buffered someplace, so tcp ends up way
overcompensating ... it backs off way too much, and now the network goes
into an "accordion" mode where it is speeding up and slowing down because
all that danged buffering isn't allowing TCP to react to real world
conditions.  Google "bufferbloat".  He hasn't done the world's best job in
actually communicating it in a way most people can understand, but he is
dead on the money.  Bigger buffers is actually probably a bad thing.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://puck.nether.net/pipermail/foundry-nsp/attachments/20110313/c2576aa0/attachment.html>