Many VGAs Assigned to One VM

[fasbit]

I have multiple VGA cards that are used for data mining (these are not on pve), they are the work-horse of the operation. I understand that I can use "passthrough" to assign a specific pci card to a specific VM, but can I assign 50 cards from 10 mobos to one VM? I need the cards to all be available to one client for easy management.

 

[m.ardito]

I guess "passthrough" does not mean "network access"... perhaps you need a network distributed engine to use 50 cards from 10 motherboards...

Marco

 

[fasbit]

I guess "passthrough" does not mean "network access"... perhaps you need a network distributed engine to use 50 cards from 10 motherboards...

Marco

Im not sure what you mean... but I need the VGAs to be accessable only to 1 VM... I guess that technically speaking they would do that through the cluster network... if thats what you mean.... Im not sure what a "network distributed engine" is either....

Maybe my real question is this: How many VGAs can one (1) PVE VM access?

 

[m.ardito]

I mean if you have a video card on a separate motherboard... how can your vm reach that by passthrough???

here
http://pve.proxmox.com/wiki/Pci_passthrough
if it helps you, is said
<<NOTE: Card support might be limited to 2 or 3 devices.>> but I don't know what is limiting the number. kvm itself?

for "network distributed engine" I mean something like a network rendering engine for 3D rendering programs,

see pages like http://en.wikipedia.org/wiki/Parallel_rendering

a bigger and differently aimed example: http://en.wikipedia.org/wiki/SETI@home

ie: your "computing" engine on the vm should (and know how to) ask other newtworked machines to do part of the work and send back results.

Marco

 

[wahmed]

@fasbit, I am assuming you are doing Bitcoin mining. A VM Reaching GPU across multiple physical hardware is just not going to happen.... yet. What you are thinking of is Parallel/Distributed computing using Beowulf for example. There are examples of Parallel Computing using CUDA, but you will need extensive knowledge on programming and time to do some Research on it. A VM can only access the hardware of the node it is in. Data mining using VM is just not possible using GPU at this moment, since KVM cannot directly acess GPU as it is.

Sent from my ASUS Transformer Pad TF700T using Tapatalk

 

[fasbit]

Yes...I am mining very similar to bitcoin mining... OK..so let me see if I got this straight... if i were to create a VM(s) that also corresponded to a physical mobo and its pci components, then and only then can the VM(s) access the pci hardware? And the way I connect these componets is through the "pass through" configuration? So really the only advantage that PVE could give me in this case would be remote manageability through the console window? Correct?

 

[Dragoon]

correct.

 

[fasbit]

So... i spent about 4 hours trying to get the VM to see my hardware with "passthrough"...as mentioned by m.ardito above... to no luck...
PVE sees them using "lspci" but the VM doesnt...
Anyone here actually used the recommended configuration?
The way its written in the wiki I think is wrong, it just says "add hostpci0: 00:04:0" to the end of the .conf file. Under this method even PVE doesnt see them...
I added "qm set 111 -hostpci0 00:04:00" to the .conf file...and at least pve registered them fine... but neither one gets them into my VM...
Any suggestions?

 

[mir]

What do you get running these commands on the proxmox host:

dmesg | grep AMD-Vi
dmesg | grep -e DMAR -e IOMMU