The budget: get the specs right (2)

So after deciding on the platform it was time to size it to our needs. Based on a quick inventory of our serverpark we estimated 80-100 servers could be virtualized. Based on the sizing guide in the excellent book “ESX Server: Advanced Technical Design Guide” (link to Amazon.com) we calculated at 4-6 VM’s per CPU core.

Going the safe route this came out as:

100VM’s / 4 = 25 CPU cores. 25 / 8 (cores per server) = 3,125 servers.

So going the safe route all the way we were budgetting for 4 HP DL585’s. Deciding on the specs of the servers itself was relatively easy after that.

We have had some questions about using Blades but at the moment they are not an option because of the VM density we are running. As described in the VMWare guidelines you need:

• a seperate NIC for the VirtualCenter management traffic
• a seperate NIC for VMotion traffic
• a seperate Gbit NIC for every 4 to 8 VM’s

With roughly 25 VM’s on a host machine that calculates to at least 5 NIC’s in the machine and the first Blades with four NIC’s have only just arrived. If you have the time and the facilities to test it all out it could work for you but as we weren’t using Blades anyway the decision was made pretty quickly to dump that idea.

The serverspecs are described in a seperate page together with some more information about how we arrived at that specific configuration.

The budget: get the specs right

Explaining the expense for the necessary hardware for a virtualization project can be a bit tricky because of the “supersizing” that you need to do. In our situation it meant explaining the difference between buying a regular two-way Intel Xeon based IBM XSeries 346 or HP Proliant DL380 (which for us do about € 5000 to € 7000 depending on specs) and buying a set of four way AMD Opteron based HP Proliant DL 585’s (which cost roughly € 25000 a piece in the specification that we need).

It all begins with getting the specifications of what you need right.

What we did was ask alot and listen carefully. We already knew how to size the systems that we have for their specific tasks so based on that knowledge we started asking other companies also thinking about virtualization how they were sizing their environment and we also started looking carefully at the current utilization of our servers.

One conclusion that stood out (at Cisco as well as at Capgemini and from our own internal VMWare consultant) was that everyone was looking at or already working with AMD Opteron CPU’s. Which made our life a whole lot easier because only* HP and Sun currently have models with a four-way Opteron configuration in their portfolio (*there are other suppliers, but because of purchasing policies I had to leave them out of the comparison)

Because we used HP already the choice was clear: Proliants DL585 it had to be. Apart from the advantage in horsepower that the Opteron has over the Intel Xeon it is also much more energy efficient and a lot less noisy. Anecdote: Visiting a trade show we could have a normal conversation with a sales rep from a server retailer standing right next to an open Opteron box. We regularly have Xeon boxes open in our room and a normal conversation is pretty impossible then.