Thanks Level1! Threadripper PCI Passthrough works smooth with Asus Prime X399-A

Hey Guys,
Just created an account to thank you for putting all the info together here. So while I am writing this from Fedora,I am listening my playlist from Windows machine. My specs:

Board: Prime x399-a
Cpu: TR 1950x
Mem: G-Skill 3200 4x8 quad channel
Gpus:
Host: Saphire toxic R9 280x, yep non uefi
Guest: Fury-x with UEFI BIOS
Disks: 2x Samsung 250G nvme (1 host , 1 guest)

Userbenchmark from running guest Here .

Crystal bench:

Host running with Numa activated and all guest devices on same node, so far pinning works no throttle:

Hardest part was splitting host and guest nvme, they had same id. Only issue I have now I can not set cpu model as host-passthrough, windows shows kernel panic, so I am on opteron g3, not even sure if it will make difference.

I am using java fix but in last 2 VM start I forget it running java app, didnt have any issue.

Thanks again & have nice day&afternoon or whatever timezone u are
Burak

9 Likes

The latest QEMU supports epyc as the cpu model. I suggest trying that.

The difference is that QEMU might not tell the OS that it supports all the fancy new CPU features (which I don’t know enough about) and may need to translate some instructions. It’s not a serious problem unless you’re doing stuff where latency is important.

3 Likes

Not to threadjack but I have found that very recently on new kernels and versions of cpufreq that they can mess with or undo overclocks somehow. I thought my tr system was feeling sluggish under load, and sure enough it was backing off to basically stock even though I’d overclocked it to 4.1ghz.

To fix I had to disable my pstate/cstate oc and go back to fixed voltage and multiplier. Extra heat and electromigration I guess. Will deal with it later.

6 Likes

Thanks I will definitely give a shot, at least I will compare g3 vs g5 vs host, what especially I wonder is hyperthreading behaviour.

First of all I am a newbie overclocker! Only thing I had similar experience with: perf dmesg messages (perf interrupt took too long) . When I had this messages, I assume I have bad V setup or memory. After increasing V they are gone. And I hope when lscpu says its 4 ghz it is 4:) I also use fixed V and Multiplier. Only addition on bios I have disabled below options:
Overlocking enhancement
Core Performance Boost
EPU power saving

increased temp a bit but when I run this machine I game or use some VM, so I dont need powersave.

On OS I run cpupower with performance governor, when I stress tested I disabled perf to catch throttling:
echo 0 > /proc/sys/kernel/perf_cpu_time_max_percent

Yet still I am looking reliable bencmarking ways, couldnt find good linux benchmarking article/material or I am lazy enough, atm just using mprime and my own script.

Hyperthreading probably requires a patch.

Phoronix testing suite. It’s the best around.

Alright it turns out that host-passthrough error was win10 v1803 issue on KVM, adding ignore_msrs as stated here https://www.reddit.com/r/VFIO/comments/8gdbnm/ryzen_2700_system_thread_exception_not_handled/dyjieg2 fixed.

4 Likes

Awesome thanks for reporting back. I had that problem myself not long ago.

1 Like

I had this issue also and added that line but as I was mentioning to @SgtAwesomesauce before, using Opteron_G2 seemed to fix things (epyc didn’t seem to work for me but I’d need to test it again) instead of copying host cpu configuration as well. A lot of guides out there say to copy host configuration, probably an issue on zen though.

1 Like

I’ve been failing, on and off, to get PCI passthrough working with Fedora on my Prime X399-A motherboard since the end of 2017. Please point me to the instructions that you used. Which version of the BIOS are you running? Sorry for the necro.

A little hope goes a long way. I finally got PCI passthrough working today. Optimization will be next.

1 Like