Statistically, most servers have low overall CPU usage (< 25%), which is the main reason why server virtualization is so effective and used for server consolidation. But you have also to consider the peaks, which could be near 100% in some cases and, depending on how frequent they are, how long in duration, and how they are correlated, could become critical for the virtual infrastructure sizing.
CPU is usually overcommitted and depending on the workload, a 4:1 ratio between virtual and physical resource could be a reasonable target if you don't have CPU-intensive (or CPU-bound) workloads. Note that the ratio should be calculated on the physical (real) cores, not the logical processors that VMware ESXi sees, because hyper-threading could increase this number. It could be also important to define how the workload could be scale with more cores or socket and how could be dependent by the architecture (not all applications perform the same on different types of processor family) or simply the speed of the processor (some old applications just require GHz instead of cores).
For the CPU, the following table suggests the main metrics to be considered:
|
Metric |
Description |
Why it's important |
|
% processor time (average and peaks) |
Percentage of elapsed time that the processor spends to execute a non-idle thread. |
Could be used to define the number of sockets and cores on the target VM, considering also a minimum hypervisor overhead. |
|
Processor queue length (average and peaks) |
Number of threads in the processor queue but not currently able to use the processor resource. |
This counter helps us to understand how many logical processors could be needed or whether there is a possible processor bottleneck. |
Finally, remember to compare apples with apples; a different processor may perform differently, and an application may perform differently by simply changing the type of processors.