VM with physical disk attached. Much slower than physical machine...

[bartplessers]

Hi,

For a migration, I have created a VM as described in https://pve.proxmox.com/wiki/Windows_10_guest_best_practices
Than I have attached a physical SSD-disk and configured passtrough, according to https://pve.proxmox.com/wiki/Passthrough_Physical_Disk_to_Virtual_Machine_(VM)

qm set 114 -scsi2 /dev/disk/by-id/ata-LITEON_LCH-256V2S-11_2.5_7mm_256GB_TW066GD55508561SH02B

On the other hand, I have my old physical machine with (another) phycical SSD attached

Both machines have same number of 4 CPU and 8GB RAM.
The physical machine has an i5-3570 CPU@3.40GHZ
The virtual machine has is hosted on a 4 x Intel(R) Core(TM) i5-7500 CPU @ 3.40GHz

I compared both machines when doing an index of my music library: I started a tool "MP3Tag" and connected to my music library (a copy of my library exists on both SSD's)

Now I noticed that my physical machine is much faster (5 a 6 times faster) than my virtual machine in doing that job.
Of course the comparison is not ideal: machines and SSD's are different, but still, the difference is huge...

I want to get rid of my physical machine, but I can not detach the internal disk for a fairer comparison.

So my question(s):
- is this behaviour normal? Do you really loose so many resources with passtrough disks?
- how can I optimize my settings?

Kind regards,

Bart Plessers

 

[Dunuin]

You can't compare different SSD models. A good one can be 1000x faster than a bad one. And qm set 114 -scsi2 /dev/disk/by-id/ata-LITEON_LCH-256V2S-11_2.5_7mm_256GB_TW066GD55508561SH02B is no real passthrough where your GuestOS can directly access the physical disk. That disk is still virtualized so you get virtualization overhead. But this overhead shouldn`t be as bad as factor 5-6.

 

[bartplessers]

You can't compare different SSD models. A good one can be 1000x faster than a bad one. And qm set 114 -scsi2 /dev/disk/by-id/ata-LITEON_LCH-256V2S-11_2.5_7mm_256GB_TW066GD55508561SH02B is no real passthrough where your GuestOS can directly access the physical disk. That disk is still virtualized so you get virtualization overhead. But this overhead shouldn`t be as bad as factor 5-6.

Hi @Dunuin

Thanx for prompt reply.

Is there another way to directly access the physical disk? How?

If not, what is a ‘normal’ overhead?

And are you aware of a more scientific way to test the disk performance? My OS is Windows10. Maybe I can shutdown my old physical server and attach it’s disk to my PVE to do a better comparison.

Kind regards,
Bart

 

[bartplessers]

Hi @Dunuin,

I just did following test

winsat disk -drive D

Results:

VIrtual Machine

> Disk  Random 16.0 Read                       221.67 MB/s          7.9
> Disk  Sequential 64.0 Read                   509.51 MB/s          8.1
> Disk  Sequential 64.0 Write                  277.99 MB/s          7.6
> Average Read Time with Sequential Writes     0.553 ms          7.9
> Latency: 95th Percentile                     2.847 ms          7.1
> Latency: Maximum                             3.707 ms          8.6
> Average Read Time with Random Writes         0.574 ms          8.6

Psyical Machine

> Disk  Random 16.0 Read                       1.39 MB/s          3.8
> Disk  Sequential 64.0 Read                   142.09 MB/s          7.1
> Disk  Sequential 64.0 Write                  100.74 MB/s          6.6
> Average Read Time with Sequential Writes     2.379 ms          6.8
> Latency: 95th Percentile                     11.978 ms          5.4
> Latency: Maximum                             44.676 ms          7.9
> Average Read Time with Random Writes         5.862 ms          5.7

So it seems that at least this test suggests that my VM is much faster!
But reading 30.000 mp3-files in MP3tag takes 6 times more time in the VM compared to the physical PC

Does anybody know what can cause this behaviour?

bart