[TriLUG] data center class x86 systems

Greg Cox glcox at pobox.com
Wed May 2 18:54:17 EDT 2012


On Wed, May 2, 2012 at 1:08 PM, Boggs, Ted
<tboggs at mcclatchyinteractive.com> wrote:
> I never understood the cabling argument from Cisco.

Incremental cable add of an 8-blade UCS (with enough cabling to cover HA):
6 U, 4 power, 4 10GbE.
Incremental cable add of 8 pizza boxes (same caveat):
8 U, 16 power, 16 10GbE, 16 FC, 1 cat5 serial.

It's so dense you can't extract a dead PS from behind the copper
without taking the box out of service.
It's so dense that you can run out of PDU receptacles before you run out of U.
It's so dense that your pizza boxes can cook because they can't push
air through the wad of cabling.
It's really an amazing visual.  Wish I could share photos.  There's
probably propaganda around for 'em.

> Most of the other major blade vendors started offering converged adapters before Cisco did.

Yup, no argument there.  CNA's are nice, but, hardly the biggest win
for UCS.  It's kinda "that's great, what else have you got for me."

> The management piece is very compelling though as long as your environment isn't too big.
> One size definitely fits most in their case.

Big is actually where it wins (and I'd say big is anything over 8
blades).  Up to 160 blades with one point of management (one place to
hit for OOB, one place to do SNMP on all the drives and cooling and
other stuff).

The real win is the Service Profiles.  Server blows up?  Tell any new
blade "you're the guy who just blew up, get in there!"  10 minutes
later, he's got the same MACs, NICs, WWPNs, obscure BIOS rules that
that wacky one-off blade needed that the last guy never documented,
and booting back into service as if he were the old guy.  Your
overnight runbook is crazy short and can be handled by the most junior
of admins.

> FWIW: You can accomplish much the same thing with tools like cobbler, koan,
> puppet, etc. if you are willing to put in a little effort up front.

Sure, if you have a unix-y stack up.  I thought we were talking more
about the hardware (pre-OS-has-booted levels), which, IMO, has always
been the thing those tools fail at.  Anyway, from a hardware
perspective, standalone rackables are dead to me.  Classic blade
chassis, too.

-G "no CSCO sponsorship expressed or implied or actual" C



More information about the TriLUG mailing list