update Nov 9th 2019
Feels like déjà vu all over again… In this NEW series of posts I’ll walk you through the steps, issues and frustrations of designing, building and configuring a low-power performance virtualization cluster based on three nodes. It’s a write-as-I-go story which means I’ll cleanup the text when I can find the time 🙂
The cluster is designed around the following principles:
- Low power consumption and low TDP
- Be as small as possible
- Silent running
- Enable high availability for VM’s
- Based on current hypervisor technology: VMWare vSphere 6.7
- Must be easily scalable and feel “enterprisy” 🙂
- Low cost solution!
As this is a complete overhaul, the principles still apply although they overall power usage is expected to be higher than the current cluster due to a different CPU/mainboard and the addition of a 10Gbit switch.
Low power consumption
To achieve this I’m using one of the latest Supermicro mainboards, the X11SDV-4C-TP8F (link) which has a System-on-Chip solution using an Intel Xeon D-2123IT CPU (link). This CPU has 4 cores/8 threads running on a base frequency of 2.20 Ghz. It is the low power version with a TDP of 60W. A passive cooling block is mounted on the chip. We’ll see if it is sufficient. I’ve been looking into other cooling solutions but the FCBGA2518 socket makes it next to impossible to find off-the-shelf watercooling for instance.
I’m not putting a drive in the nodes but I’m going to boot from my NAS. My existing QNAP TVS871 will provide the necessary iSCSI LUN’s. The NAS is always on anyway so this saves on power consumption for each node.
Be as small as possible
Unfortunately I had to move away from the mini-ITX form factor which means my existing cases for the cluster cannot be used anymore. So, I’ve decided to move to an 1U Supermicro SuperChassis 505-203B case (link). The chassis is different than ‘normal’ 19″ versions as the mainboard is mounted rotated 180 degrees which means all the connections will be on the front panel and easily accessible when mounted.
I’m moving everything into a spiffy new small 12U sized 19″ wallmounted rack, the Caymon OPR312A (link). This rack is perfect as it is resizable in depth. All of my components are relatively short so, at most, the rack will protrude 300mm from the wall. It can hold a load of about 50kg which is plenty for this build.
“As small as possible” also means no additional hardware other than the clusternodes.
Silent running
Even though I have a separate room for my hardware, it needs to be as quiet as possible. The passive cooling of the Supermicro boards will probably not be enough so I’m adding high quality quiet cooling fans to each chassis, the Noctua NF-A4x20 (link). Most 1U server boxes are very noisy due to the small fans needing to have a high airflow meaning high RPM’s. The Noctua’s special design will provide about 9.4 m3/h while still producing only 14,9 dB(A).
Enable high availability for VM’s
A three-node-cluster with the ability to have zero downtime for VM’s which live in the cluster
VMWare
The virtual layer is based on the spiffy new VMWare vSphere 6.7 platform meaning the nodes will run on ESXi 6.7 and management will be done through vCenter Server 6.7 appliance.
Must be easily scalable and feel “enterprisy” 🙂
It must be easy to add more nodes without going through complex design/configuration stuff.
Low cost solution!
Speaks for itself. Well, it ‘did’ previously. This build however will be a bit more expensive as I’m only reusing the storage and MER switch. It is still a low-cost build looking at the specification and value-for-money but it is much more expensive than the previous clusters.