First io-smp-test with kernel 2.6.32-1-pve

[udo]

Hi,
i have just installed the newest kernel and make some io-test with a win-kvm vm and h2benchw.exe.
Unfortunallity the index is very low (app. 2) if the vm has 2 cpus or 1 cpu with 2 cores assigned. With one cpu the same index as under pve-1.4 are reached (app. 84).
Under pve-1.4 the index varied between 6 and 34 with 2 cpus.

Hmm, i will try tonight a rip-process with the new kernel and 2 cpus and will post the results soon.

Udo

 

[tom]

Hi,
i have just installed the newest kernel and make some io-test with a win-kvm vm and h2benchw.exe.
Unfortunallity the index is very low (app. 2) if the vm has 2 cpus or 1 cpu with 2 cores assigned. With one cpu the same index as under pve-1.4 are reached (app. 84).
Under pve-1.4 the index varied between 6 and 34 with 2 cpus.

Hmm, i will try tonight a rip-process with the new kernel and 2 cpus and will post the results soon.

Udo

pls try also 2.6.18 to see the difference.

 

[udo]

Hi,
for the rip-test i got not the time yesterday - i think this evening can i do it.

I do the h2benchw also with 2.6.18 and the results nearly the same. Sometime the index rise up to 10-13, or it's around 2. No different between 2 cpus or 2 cores.
With this test the vm got not more than 50% cpu. If one cpu works, the other idled. If the second rise, the first go down.

Udo

 

[udo]

Hi,
i made the rip-test ( described before in
http://http://www.proxmox.com/forum/showthread.php?t=2731 ).

Now the result looks better!
Here comes the times to rip a 950MB-File:

pve1.3 2CPUs 3:45
pve1.4 2CPUs 5:07
pve1.5 2.6.18 2CPUs 4:04
pve1.5 2.6.32 2CPUs 2:52
pve1.5 2.6.32 4CPUs 3:30 -> this shows that the scaling could be much better.
But i reach my goal to be significant faster with proxmox - but only with 2 CPUs.

To remind, the same process with the "big player"-vm takes (on slower hardware) 4:20. If, for comparing only the bogomips are taken, pve1.5 2.6.32 2CPUs should need 3:35 on the different hardware - but this is not comparable.

Udo

 

[udo]

Hi,
i made the rip-test ( described before in
http://http://www.proxmox.com/forum/showthread.php?t=2731 ).

Now the result looks better!
Here comes the times to rip a 950MB-File:

pve1.3 2CPUs 3:45
pve1.4 2CPUs 5:07
pve1.5 2.6.18 2CPUs 4:04
pve1.5 2.6.32 2CPUs 2:52
pve1.5 2.6.32 4CPUs 3:30 -> this shows that the scaling could be much better.
But i reach my goal to be significant faster with proxmox - but only with 2 CPUs.

To remind, the same process with the "big player"-vm takes (on slower hardware) 4:20. If, for comparing only the bogomips are taken, pve1.5 2.6.32 2CPUs should need 3:35 on the different hardware - but this is not comparable.

Udo

Hi,
now we have sucessfull move the rip to the new proxmox-server (pve 1.5 with 2.6.32) and it run's fast and stable.
But one curios things happens: I move the rip-server from the test-machine (where a big rip-job takes 2:52) to the new server (with faster CPU and faster IO). I made the same rip-job and it takes 3:20!! What's happens? Normaly all things are faster, but the job takes longer. Then i switched to only one CPU. The rip process takes longer, but faster as under the same condition on the test-server (slower IO+CPU) - that's what i expected. SMP works not very reliably...
And i have guranteed the top-performance. So, i switch the Rip-Server from IDE to virtio and i got with 2 CPUs again the 2:52... Thats ok, but i think there is much more possible with better working SMP.

Udo

 

[ellenverheyen]

Which exact command did you use to test your hd?

 

[udo]

Which exact command did you use to test your hd?

Hi Ellen,
in this case it's a special software to rip (raster image processor) postscript-data to a bitmap-file. It's needs good cpu-performance and also io-performance (read the file and write to another).

You can test the pure io-performance (windows) with h2benchw - see post http://www.proxmox.com/forum/showthread.php?t=2731
This program show, that the io-performance with one cpu is good, but with more cpus not.

Udo