[Updates below w/ new, more specific requirements!]
Does anyone know of any source at all of benchmarks looking at shared-graphics/API-redirect in Libvirt-Qemu-KVM or Android-Studio (or any other hypervisors)?
Specifically, looking at; max. number of active guests supported while maintaining host system responsiveness, lowest overall power consumption, well supported graphics tech., hardware availability if possible.
If not, are there any sites/channels you would point to as possibly being interested in doing this accross a wide range of hardware generations?
Thank you!
(My core problem is in building a system [Ubuntu-Desktop-22.04] that can function primarily as an Android-Studio devstation/build-server. My current main devstation [SandyBridge: e3-1235} is completely unusable w/ more than 2-3 active clients taxing the igpu, & my laptop [IceLake: 1035g1] can’t handle more than 1 w/o turning into a leafblower.
So primarily; is there a low power graphics card that solves this problem? Would making the jump to a Ryzen APU do it?
I think I should clarify:
For hardware compatibility reasons any graphics card reccomendations must be one of the following.
[Section reposted from STH]
An important consideration for me is that all main(always-on) systems pruduce zero fan noise. The laptop has a fanless mode that is otherwise sufficient. I’ve considered removing the fan entirely, but am unsure of what would happen during heavy workloads.
If anyone has access to hardware that they think can fulfill this requirement, & would be willing to do some light testing w/ Android-Studio AVDs (or alternately; Virt-Manager, which also uses Qemu-KVM), that would be most appreciated.
I hope that made sense!)
This article makes no sense to me, please help! https://www.anandtech.com/show/2480/12
Other than restating that desktop based virtualisation-hosts are a bad idea, is there anything useful to be gleaned here?
VRAM usage is #1 concern, w/ only 5xAVDs using ~1600/2048M, while the actual GPU on-die resource usage remains comically low throughout.
I would think performance-per-watt is probably the next thing down the list, w/ the SandyBridge architecture clearly struggle-juggleing to keep up w/ too many simultaneous guest processes.
CPU utilization is highest (40-100%) during AVD boot sequences, & then drops down to ~5-8%, while main memory usage is actually quite low now, at 2020M(used)+9670M(buff) , & the system remains respopnsive w/ entire graphics load taken off IGPU.
FurtherResearch:
S4000x is already at the edge of what this configuration can probably handle, so probably don’t bother w/ MXM card recommendations at this point. My strongest preference would be for the half-height, 6inch, low-power variety of graphics card, please especially shout out any of those that may fit the above requirements. Also, is there any way of predicting/tracking GPU performance degradation, going from GDDR5-to-DDR3/4 graphics cards, or even main memory (& are there any of those w/ an unnecessarily large framebuffer)?
I misconfigured my LAN somehow, & will have to figure that one out before I can get some hardware video accelleration numbers.
Will also have a hunt around for some good test activities. Any suggestions welcome!
Paravirtualisation still causing occasional crashes & kernel-panics, as expected. would love to know if anyone using newer architectures still has this problem, or if modern VTX/VTD/GVT extensions have solved the issue. Also, the current situation w/ AM4s SVM extensions on Ryzen 1st-4th gen (& Athlon 3xxx maybe) is of course, a matter of the highest interest for me.