For example, we found two orders of magnitude performance difference between PV and HVM on ec2 on fork-heavy CI jobs. A difference between 3 and 30 minutes. This difference is trivially reproducible by measuring the time taken to complete a fork() syscall. As the memory size of the parent process grows, the difference becomes more pronounced.
Here's actual data I collected from c3.large instances, using a contrived Ruby one-liner to benchmark fork():
PV: [ec2-user@PV ~]$ ruby -e'$ref=[]; (24..29).each {|n| $ref << "X" * (2 n); start = Time.now.to_f; fork.nil? and exit!; puts Time.now.to_f - start; sleep 1}'
0.00550532341003418
0.011992692947387695
0.02542281150817871
0.05213499069213867
0.10562920570373535
0.21283364295959473
HVM: [ec2-user@HVM ~]$ ruby -e'$ref=[]; (24..29).each {|n| $ref << "X" * (2 n); start = Time.now.to_f; fork.nil? and exit!; puts Time.now.to_f - start; sleep 1}'
0.0005323886871337891
0.0007219314575195312
0.0012335777282714844
0.002183198928833008
0.004210233688354492
0.008102178573608398
As you can see, the difference is considerable. This is all on Amazon ec2, the only difference between these VMs is PV versus HVM.
(edit: HN formatting seems to be eating the exponent operator in the "2 to the n" expression. It's a double *)
also note that fork() involvs "lots of" io computations which could also be a driver problem and they evolved in newer kernels, too. and also fork() could syntactially be different on hosts with ballooning and hosts without. especially on kvm.
edit: as of today we switched to hvm instances (however we are on many many micros so that performance boost is likewise zero) on our inhouse hardware we are varying between kvm and vmware. however I dislike vmware more and more. vcenter is just unusable (flash) and consumes a shitload of memory. so mostly we use vmware api calls which makes vmware unnecessary somewhat. also we sell our software as a appliance with a xenserver host and mostly we tested our software an all ends and we are mostly disk io driven and that barely changes between them. there are some aspects that could improve performance but that depends more on how you configured your virtual disk.
32bit, or 64bit pre-virtualization hardware extension (vt-x/d, ept, amd equivalents) are really the only ways this should have been the case. Each time you make a fork() system call on 64bit PV you're making an extra context switch and flushing the TLB that an HVM instance does not need
I would expect a full description of the benchmarks and setup to follow up.
After reading your comment, and admittedly not knowing much about hypervisors (yet), I am genuinely curious as to what else could be causing this drastic difference in performance, if not for the hypervisor itself. Any ideas?
Real hypervisors are utilising a lot of hardware features to accomplish jailing. Client hypervisors may or may not be, depending on how they're implemented, and or how they're integrated into the underlying OS (e.g. specialised drivers, just simulating hardware features, etc).
You should be looking at dedicated virtualization hosts to do an apples to apples comparison, like vmware on bare server metal vs hyperv or any of the free tools also on metal. Not on a cpu throttled desktop.
I would always separate benchmarks for server or desktops.
The difference between hypervisors for IO is tiny. They're all slower than native, but vary very little between one another. That's why the manufacturers have all almost given up trying to use performance for comparison, and instead argue value adds/management/automation/cost.
If you really found one specific hypervisor slower than another (regardless of which two), I'd want to look at your exact test conditions since that has never been the case for me (everything else being equal). And even the manufacturers don't argue that point much.
There are plenty of reasons to e.g. pick AWS over Azure, but Hyper-V performance is not one of them. That's just FUD.