Doesn't Your New Host Deserve a High Performance Hypervisor?


Yes, VMware used to be the heavy hitter.

Following on a little from "Doesn't Your Server Deserve a High Performance Operating System?", it's worth comparing performance on a layer some might regard as pre-operating system:  the hypervisor.

In many IT shops, the hypervisor is actually in control of your hardware, and therefore has a strong influence on how much computational capacity you're going to realize from your investment.  But not all hypervisors are created equal. 

Sadly, there's not quite the sport of drag-racing hypervisors the way the Top 500 supercomputers roil out smoke and thunder.  Nevertheless, some intrepid folk do compare hypervisors:

http://spec.org/virt_sc2013/results/specvirt_sc2013_perf.html

Not many new contestants since Linux KVM spun up a modern engine.  Where is XenServer?  Hyper-V?  Oracle VM?  Frightened, I'll warrant.

But, delightfully (for the number-crunchers amongst us) there is an apples-to-apples contest where RHEL6.4's KVM gets to compare freight-moving power against VMware ESXi 5.1 - using the exact same Intel processors, the same amount of RAM, hypervisor on SAS, VMs on SSD.  The results, in number of guests handled?  KVM:  37.  ESX:  27.  Oooh, how the mighty have fallen. 

Let me go for an aside, right here.  There are those hypervisor adherents out there that will tell you that VM performance is 'essentially native' or 'near-native' in comparison to on-the-metal performance.  But, if you go and talk to e.g. a DBA, they might tell you that your database server should go in a VM 'over my dead body'.  Who is right?  I think SPECvirt is telling you, right here.  If KVM can produce a 37:27 performance ratio over ESX, then ESX couldn't have been 'near-native.'  At best, if KVM was 'near-native', then ESX was only giving you 73% of your hardware.  And I'll bet that KVM isn't really 'near-native' either.  The upshot?  Implement hypervisors with a mental caveat that you're eating a chunk out of your infrastructure's capacity before you start any work.  Hypervisor virtualization is inefficient.  Aside completed, back to KVM vs ESX.

If your environment needed 100 such VMs (SPECvirt uses a balance of application, web, mail and batch VMs), KVM could do it in 3 servers, while ESX would need 4.  That's a business case.  You don't even need to look at KVM's zero-licensing-cost and lower support fees to recognize that ESX is bad value for money.

So, leave your romantic notions of days gone by for some other poor sap.  Let them spend money on canvas and teak and brasswork.  You'll be better off with turbines and steel and RADAR.

So, if you're dead-set on hypervisor virtualization, doesn't your new host at least deserve a high performance hypervisor?