Our Virtualization Management Index showed some interesting data on the average size hardware enterprises are using in their datacenters to support VMware. The average system in the study contained:

• 2.4 sockets per host
• 3.6 cores per socket
• 1.8TB of storage
• 50GB of memory
• CPUs running at 2.6 GHz.

This data got me wondering - what is the "ideal" platform for VMware? Is there an ideal platform? How do enterprises make this decision? Scale up or scale out?

I looked at five different combinations of hardware. Costs listed below include the ESX license.

• Dell R610 - 2x2, 96 GB of memory. $ 20,415.
• Dell R610 - 2 x 2, 192 GB of memory - In this configuration you pay a penalty over the 92GB configuration for using 16GB DIMMS since they are more than twice the cost of 8GB DIMMS. $ 31,597
• Dell R610 - 2 x 2 running at faster processor speed of 3.4 GHz, 96 GB of memory - In this configuration, you can't get higher memory density over the slower running system since the higher processor speed does not support 16 GB DIMMs. $ 27,784
• Dell R900 - 4 x 4, 128 GB of memory. $ 50,199
• Dell R900 - 4 x 6, 256 GB of memory - this configuration is a monster! The six core configuration also adds more costs on a cost/core basis than the 4 socket configuration. $ 75,390

Which is the correct hardware platform for a VM environment?

To figure this out, I made the assumption that to find the right size hardware platform, I needed to first understand total CPU and memory requirements for my entire environment.By adding this up for every VM, assuming peak values, this would provide the total CPU and memory peak capacity I required.The next step would be to find the cheapest combination of server types to meet this CPU and memory need.

To perform this calculation, I need to make some assumptions for my sample environment. The assumptions are:

• average peak values for virtual machines - 1.2 GHz and 1 GB of memory
• an environment of 1,000 VMs
• cluster size of 5 hosts with 25% reserve failover capacity

Which configuration was the winner for our theoretical environment?

The Dell R610 (2 x 2 Fast -96) running at the faster processor speed but with less memory is the best fit for a majority of combinations of VM size.  All the blue area in the graph is where the 2 x2 Fast 96 has the best ROI.

Special note here is that in this case, the winning configuration is based on measuring actual peak values for the virtual machines. This is almost never the case in sizing exercises (unless you are a VKernel customer of Optimiztion Pack) as most virtualization administrators will use sizing based on physical world allocations or guesstimates based on standard VM sizing. This could result in disastrous oversizing and selection of the least efficient platform.

Incorrectly calculating required CPU and memory resources can have a drastic effect on server selection.Note in the graph above as the required memory per VM increases, the optimal server actually changes from the 2 x 2 FAST CPU configuration to the monster 4 x6 configuration.  Of course this may also introduce concerns about putting too many VMs on a single machine.  But theoretically, the optimal machine size shifts.

What is the impact of making an improper selection of core hardware? The cost impacts can be disastrous. The chart belows show the delta between the optimal server platform cost and the most expensive platform for a given configuration. Millions of dollars can be overspent with the incorrect selection.

What is the net of this?

1. Ideally, understand peak average CPU and memory configurations for your entire environment based on actual usage.
2. Find the most optimal platform to supply these CPU and memory requirements at a VM host density you are comfortable with.
3. A mistake in hardware selection can cost millions
4. Constantly re-evaluate based on memory and CPU economics
5. The "cheapest" server may not be the best just like the server with maximum memory may also not be ideal. It all depends on the average peak loads.
6. And of course, use VKernel Optimization Pack to actually understand peak loads in the environment to drive down the cost of hardware.

Bryan Semple
CMO
VKernel