
Hi,
On 5 Nov 2014, at 2:15 pm, Peter Ross <Petros.Listig@fdrive.com.au> wrote:
Maybe a benchmark question: Which benchmarks would you run to compare the two in various ways?
Depends entirely on the workload you're running in the guest. :) For Oracle RDBMS for example, I'd run things like Swingbench, ORION, SLOB and other DB-specific benchmarks. I'm not sure what the most appropriate benchmark for a Zimbra box would be. Usually I do not over-allocate (sorry, there is a better word I just
cannot remember just now) or at least not much. The space is also used as a buffer if I have to migrate VMs/services suddenly (e.g. caused by a failing machine)
I suspect the phrase you're looking for is over-subscribe, though over-allocate makes perfect sense too. In the Oracle VM space, we always tell customers to have at least 1 server's worth of VM-allocated resources free. This can sometimes be challenging. If every server has 64GB of RAM, and you have a VM with 32GB of RAM, you need to have a server with over half its memory unallocated in order to migrate that VM. Another reason why more, smaller VMs are more agile. :)
The result may differ on different hardware as well. E.g. I could imagine to have different results if I try the two stacks on a 32GB modern hardware - or on a 5 years old machine with 8 GB only. Depends where I find a bottleneck.
Exactly. And these days, hyperthreading makes a difference too. Everyone things they want "double the processors", but depending on workload, NUMA architecture and hypervisor, hyperthreads can actually cause CPU spin bottlenecks. You should always benchmark your system with and without them enabled to see if it makes a difference for your particular workload. Cheers, Avi