Fact: Everyone’s going virtual to reduce cost and improve utilization. Why not? Sharing of resources and paying for resources on-demand should be beneficial. But that just makes the job of the capacity planner even more difficult!
3 big challenges in Virtualization that do not exist in traditional capacity planning:
1. Capacity of each box is dynamically allocated, so which virtual machine (VM) actually got how much of resources?
2. Each box has overhead utilization from each VM due to the hypervisor which reduces performance, so what is the critical number of VMs to be configured on each box?
3. Cost models are complex with options for on-demand vs dedicated vs burst capacity, so which option should be chosen?
Solutions: 1. We use a couple of key performance metrics (Transaction Rate and Response Time) that are uniform across all layers and applications. This tells us how much capacity was effectively used by each application.
2. We use a slowdown factor that discounts available resources due to hypervisor utilization. As a result, we can derive the optimal number of VMs on each box.
3. Because of solutions 1 & 2, we are able to accurately forecast capacity requirements which can then be compared with the different cost models. If capacity requirements are high but stable, dedicated would be cheaper, but on-demand is better for unstable capacity requirements.
3 big challenges in Virtualization that do not exist in traditional capacity planning:1. Capacity of each box is dynamically allocated, so which virtual machine (VM) actually got how much of resources?
2. Each box has overhead utilization from each VM due to the hypervisor which reduces performance, so what is the critical number of VMs to be configured on each box?
3. Cost models are complex with options for on-demand vs dedicated vs burst capacity, so which option should be chosen?
Solutions: 1. We use a couple of key performance metrics (Transaction Rate and Response Time) that are uniform across all layers and applications. This tells us how much capacity was effectively used by each application.
2. We use a slowdown factor that discounts available resources due to hypervisor utilization. As a result, we can derive the optimal number of VMs on each box.
3. Because of solutions 1 & 2, we are able to accurately forecast capacity requirements which can then be compared with the different cost models. If capacity requirements are high but stable, dedicated would be cheaper, but on-demand is better for unstable capacity requirements.
