Benchmarking with poor results for proxmox

[groy]

Hi,


A benchmarking on vsphere, hyper-v, proxmox and xen was published on Tom's Hardware
http://www.tomshardware.fr/articles/solutions-gratuites-virtualisation-hyperv-vsphere,2-907.html


Bad performance so bad publicity for proxmox...
The bench was made with local disks and network storage.
Tests are made with qcow2 disk, i have reported a raw disk would provide best performances. But it's doesn't explain everything.


Some feedbacks on these benchmarking by the proxmox team ?


The reporter is open to explanation on these bad results.


Bye,
groy

 

[hstende]

Running one vm in a test isn't that interesting. Its more interesting running test with more vm's to see how it scale. Like using ballooning drivers in windows etc.
Another thing that impact is how the underlying fs is made (ext3, ext4,xfs) and how its tuned.
What cpu features is allowed by the vm like -cpu flag with KVM or EVC mode wtih vphere will also impact on performance.

hste

 

[aderumier]

I check the iometer resultats
50000iops with 8 local disks raid5 ?????
12000iops with RN716X san and 6 disks raid5 ????

Seem that theses benchmark are bullshit

Or they use writeback in other hypervisors

 

[e100]

One screenshot shows IDE not virtio: http://www.tomshardware.fr/image/page12-pic12,0101-433242-0-2-3-1-png-.html

In another they have write back in use (not helpful for performance in most instances and not useable for benchmarks )

I also think that a single VM in KVM is likely not going to perform as top dog but when running many VM KVM will perform at the top.
Who sets up a virtualization system to run a single VM anyway?

I know for a fact that the aggregate IO of a few VMs can saturate the backing storage.
Can't ask for anymore performance than that from any system.

 

[e100]

Qcow vs raw...

If you actually want to compare apples to apples then using raw with no cache in all of them is the only option.

 

[mo_]

to be honest, local tomshardware sites (i.e. not .com) are known for doing bad (as in: using terrible settings, not using the same settings for the same tests and so on) reviews and such, so that really doesn't say anything, as the others have already pointed out