Lately on the Twitters there has been much praise among my friends and colleagues for what I like to think of as datacenters on dollies: Cisco’s UCS, FlexPod, Dell’s vStart etc…You know what these are as I’m sure you’ve come across them: pre-configured, pre-engineered datacenters you can roll out to the datacenter floor, align carefully and then -put your back into it lads!- carefully drop onto the elevated tiles. Then you grab that bulky L1430P and jack the stack into your 220v 30 amp circuit that has A/B power and bam! #InfrastructureGlory achieved.
Support’s not a concern because the storage vendor, the compute vendor, and the network vendor are simpatico under the terms of an MOU…you see, the vendors engineered it out so you don’t have to download and memorize the mezzanine architecture PDF. All you have to do now is turn it on and build some VMs in vSphere or VMM or what-have-you.
Where’s the fun in that?
Don’t get me wrong, I think UCS is awesome. I kind of want an old one in my lab.
But in my career, it’s always been pizza boxes. Standard 2U, 30″ deep enclosures housing drives & fans up front, two or four CPU sockets in the middle surrounded by gobs of RAM, and NICs…lots and lots of NICs guarding the rear.
And I wonder why that is. Maybe it’s just the market & space I tend to find employment in, but it seems to me that most IT organizations aren’t purchasing infrastructure in a strategic way…they don’t sit down at a table and say, ‘Right. Let’s buy some compute, storage, and network, let’s make it last five years, and then, this time five year’s from now, we’ll buy another stack. Hop to it lads!”
A good IT strategic planner would do that, but that’s not the reality in many organizations.
So I’ve come to love pizza boxes because they are almost infinitely configurable. Like so:
- Say you buy five pizza boxes in year 1 but in year 2, a branch office opens and it’s suddenly very critical to get some local infrastructure on-prem. Simple: strip a node out of your handsome 10U compute cluster and drop-ship it to the branch office. Even better: you contemplated this branch when you bought the pizza boxes and pre-built a few of them with offlined but sufficiently large direct attached storage.
- You buy a single pizza box with four sockets but only two are populated in year 1. By year three, headcount is surging and demand on the server -for whatever reason- is extraordinary. What do you do hotshot, what do you do? Easy: source some second-hand Xeons & heatsinks, drop them into the server and watch your cpu queue lengths fall (not quite in half, but still). But check your SQL licensing arrangements first and be prepared to consolidate and reduce your per-socket VMs!
- Or maybe you need to reduce your footprint in the datacenter. If you bought pizza boxes in a strategic way, you just dump the CPUs and memory out of node1 into node 2, node 3 into node 4 and so on. You won’t achieve the same level of VM density but maybe you don’t need to.
- Or maybe you don’t want or need 10GbE this year; that would require new switching. But in year 2? Break a node out and drop in some PCIe SFP+ cards and Bob’s your uncle.
I guess the thing about Pizza boxes I like the most is that they are, in reality, just big, standardized PCs. They are whatever architecture you decide you want them to be in whatever circumstances you find yourself in.
A FlexPod or vStart, in contrast, feel more constricting, even if you can break an element or two out and use it in another way. I know I’d be hesitant to break apart the UCS fabric.
You’d think a FlexPod would be perfect for small to medium enterprises, and in many cases, it is. Just not in the ones I’ve worked at, where costs are tight, strategic planning rare, and the business’ need for agility outstrips my need for convenience.
Also, isn’t it interesting that when you compute at “Google-scale” (love that term, is it still en-vogue with VARs?) or if you’re Facebook, you pick a simple & flexible architecture (in-house x86/64 pizza boxes) with very little or no shared storage at all. You pick the seemingly more primitive architecture over the highly-evolved pod architecture.
Wilson.tech 
I am indeed Sean. Thanks for the correction!
LikeLike
Great post. I think you’re confusing UCS, which are Cisco’s blade (B-series) and rackmount (C-Series) servers, with vBlock, which is Cisco and EMC.
LikeLike