KVM performance on different operating systems

[snowman66]

Yes, but the official recommendations for proxmox is still to use nocache but this is clearly only true if the images is storage on a storage with hardware raid.

Above test was done using hardware raid controller. 6disks in raid 10 with BBU.

 

[mir]

Above test was done using hardware raid controller. 6disks in raid 10 with BBU.

Obviously the recommendation needs to be adjusted then. This also follows the changes in Redhat kernels from RHEL 6.x where numerous tests and official Redhat recommendation is write-through or write-back. This is documented here: http://www.ilsistemista.net/index.p...s-on-red-hat-enterprise-linux-62.html?start=2

 

[snowman66]

I see the point. Thanks.

I did one more test because of strange I/Os and i can see i made something wrong?
If i re-run the benchmark 3-4 times in period then i get results like above on the first run results are much lower.

 

[spirit]

Yes, but the official recommendations for proxmox is still to use nocache but this is clearly only true if the images is storage on a storage with hardware raid.

if you use writethrough, you will have read cache on host, but you have already read cache in your guest buffer memory. That's why cache=none is better.
also writethrough is slower (but more secure) for write.
more info here:
http://pve.proxmox.com/wiki/Performance_Tweaks

 

[mir]

if you use writethrough, you will have read cache on host, but you have already read cache in your guest buffer memory. That's why cache=none is better.
also writethrough is slower (but more secure) for write.
more info here:
http://pve.proxmox.com/wiki/Performance_Tweaks

Have you read this? http://www.ilsistemista.net/index.p...s-on-red-hat-enterprise-linux-62.html?start=2

 

[mir]

if you use writethrough, you will have read cache on host, but you have already read cache in your guest buffer memory. That's why cache=none is better.

Every performance test I have run shows that cache=none performs worst.

 

[spirit]

Ok, I have begin to do sandra benchmarks under win2003, I can confirm that benchmarks doesn't involve disk access for the performance difference that docent see.
So it must be related to cpu or memory bandwith.
I'll do the test with differents kernel, and post the result at the end of day.

 

[spirit]

oK, i have done the sandra bench on 3 differents generation dell server, with pve-kernel and debian kernel, and I got exactly the same results.

So Docent,i think the problem must be related to your hardware config, maybe a different module driver version, or something like this.

I can't help you more than that, sorry .....

 

[docent]

oK, i have done the sandra bench on 3 differents generation dell server, with pve-kernel and debian kernel, and I got exactly the same results.

So Docent,i think the problem must be related to your hardware config, maybe a different module driver version, or something like this.

I can't help you more than that, sorry .....

Ok
Thanks
I'll look for the cause