[138204] in North American Network Operators' Group
Re: Switch with 10 Gig and GRE support in hardware.
daemon@ATHENA.MIT.EDU (Anton Kapela)
Mon Feb 28 22:07:15 2011
In-Reply-To: <13702_1298181629_p1K60NTu008696_93717E08238F7B4392FAD1F82A87FBD3365F4961@SAT2EXD02.RACKSPACE.CORP>
Date: Mon, 28 Feb 2011 21:07:06 -0600
From: Anton Kapela <tkapela@gmail.com>
To: Matt Newsom <matt.newsom@rackspace.com>
Cc: "nanog@nanog.org" <nanog@nanog.org>
Errors-To: nanog-bounces+nanog.discuss=bloom-picayune.mit.edu@nanog.org
On Sun, Feb 20, 2011 at 12:00 AM, Matt Newsom <matt.newsom@rackspace.com> wrote:
[snip]
> I can't seem to find anyone that has small 1-2U solution that can do the full shake and bake.
[snip]
If you don't need "all ports, all-line-rate, all the time" ... you
could rig it with two rack units. I'd dirty it up with something like:
First U: Brocade CER2k 48 (NI-CER-2048CX-ADVPREM-DC) 2x 10 gig + 48 sfp
Second U: Arista 7148SX (DCS-7148SX-R) -- 48x 10 gig
Ridiculous config #1231512:
-take CER 2x 10 gig XFP, configure LAG & trunk host-facing and
transit-facing vlans towards Arista 7148
-configure remaining 46 sfp+ ports as access interfaces for hosts on 7148
-configure static GRE on CER2k, transit neighbors, etc.
-remember to crank mtu's on *
-if box-router traffic is hashable, you could slum it with N x 1gig
port LAG's for more bits/sec between the 7148 and the CER2k
Of course, if two or more hosts are active, chances are good their
inside-tunnel packets won't hash to either side of the 2x10g lag from
the 7148. If your app is using IP in a more i-mix profile, the inner
traffic could then be likewise flow-dense, meaning the hashing will
effectively permit you to utilize all 20 gbits. Arista hashes on
7-tuples, whereas CER is l2+3+4 + labels (if doing mpls, etc), as of
v5.1.
I hear rumors that Brocbroc will be doing 1u N x 10 gig box where N
will hopefully be >= 8 ints... of course, that doesn't help you now.
Lastly, one could employ a 6716 blade + sup720b non-xl (for cheapness)
in 6503E/7603 chassis... but ugh -- 4u, and almost 2x the runtime
wattage of the ghetto-rigged config, and *still* internally
oversubscribed to not-quite-40 gigs per slot!
-Tk
