#Level1Presents
ECC Memory Support on Threadripper?: 6:20 Newer Kernels wont boot because of SEV/PSP: 7:51 Process Lasso and NUMA: 10:50
“You got a dummy die and a dummy die, real die and real die… We don’t know which one’s which…”
der8auer used 3M Novec fluid to find the active dies @wendell
Also, I prefer to dedicate a NUMA node strictly to a Passthrougu VM with the other node running the host.
Definitely want to see some x264 and x265 encoding tests with the threads=180 parameter.
@wendell You can get John the Ripper working on Fedora 28 by (temporarily) installing openssl version 1.0.2 instead of version 1.1.0 withsudo dnf install compat-openssl10-devel --allowerasing
I figured that out and the mpi/hpcc problem also, frustratingly. I’m using Debian on the system right now just to sanity check performance on fedora.
Same problem on Ubuntu but you can apt install oldfart and it works
Which linear algebra library did you end up going with?
Note that if ccache is install on your system, it might cause some weirdness in phronix test suite’s ability to see GCC: https://www.phoronix.com/forums/forum/phoronix/phoronix-test-suite/1044372-why-doesn-t-pts-system-info-see-gcc
Hopefully you test on Fedora too… Also, for sanity testing you can try the PTS docker image.
Openmpi and whatever was in phoronixs list of “install this”
The docker image is next on my list but I need to figure out how to update the 7zip binary on windows. It’s hilariously out of date. And I’d love it if docker would work on windows. But it doesn’t work like that.
Ok, it would be great if you leave enough breadcrumbs for reproducibility of results (exact installation configuration would be great!) .
Btw, using the docker image seems pretty painless (on linux).
Which list? This installation script lists several alternatives for MPI and LA libraries: https://openbenchmarking.org/innhold/1b19b3234f1f8f150c0ec942a6fff99d5828d0ba
I commented over on YouTube, but today’s video finally pushed me over the edge to register on here to post. I’ve been a viewer for a long time and I cannot overstate how much I enjoy your content, thanks for all the videos.
Quoting myself from YouTube:
The AGESA/SEV issue is a bit more nuanced than you’re saying. Blacklisting modules or not compiling a module isn’t a viable temporary fix due to a dependency between psp, ccp and (most importantly) kvm, so if you want to use KVM your only solution is to downgrade the bios.
Even this timeout patch is a good month or so away from being propagated to end users. Best solution right now is to just downgrade the bios, IMHO.
Edit: I can’t completely confirm this, but on a x470 Taichi Ultimate the default SEV configuration is “disabled” and I believe I hit this issue on the newer AGESA even with it disabled. So I don’t believe that’s a viable fix, at least not on every board.
I just upgraded again to test this last part more solidly, and I can definitively say that with an ASROCK X470 Taichi Ultimate and a 2700x, this SEV issue appears regardless of whether support is enabled or disabled in BIOS. This makes sense because the AMD correspondence on the RH bugzilla makes it sound like the issue is within the AGESA blob rather than a mistake that all of the MB vendors are all independently making. With that in mind, I’m actually surprised that you’ve found a motherboard where disabling SEV does fix the issue.
Containers perf comparison… uma vs numa
It’s a special bios from asrock but the fix is incomplete. I can use kvm at least but I still must disable the PSP at compile time. Or use an older kernel. Older bios isn’t a real fix since I want the PCIe reset fixes in the new agesa.
Openmpi and openblas. When running the test suite it gives a long list of packages missing, and I just installed all those without reading or selecting. I think openmpi/blas is what was in that list. I didn’t make a conscious selection. Though it did NOt autodetect mpi but did detect openblas. I updated Make.pts with the mpi paths and that worked fine. On fedora. On ubuntu it worked fine.
I think a good workload to test for a DevOps machine would just be source code compilation for larger projects as kind of a mock CI pipeline.
A good test project I use is the Dolphin emulator (https://github.com/dolphin-emu/dolphin). It’s free, cross platform, easy to compile, and the codebase is a decent size so it doesn’t finish in 2 seconds.
Another good test project is Kodi (https://github.com/xbmc/xbmc) again becauses it’s free, cross platform, relatively easy to compile, and the codebase is larger.
What exactly do you need 128GB of RAM for? Are you a NASA scientist?
I imagine that RAM kit cost you as much as high-end gaming system.
Tons of people could use that ram, Think of all the chrome tabs you could have open, but in all seriousness you could use about half of that consistently with a lot of workloads
First I want to thank asrock for being so cool! Now I feel comfortable buying fitality to my 2950X system(s), thanks. Well mostly thanks to Sir Wendell 
To my question, I wonder if it’s possible to see some tests with VM on threadripper 2, like on ESXi. Or perhaps some in chat might have some experience to share? I know it’s really early and the cpus barley released but hey, one can only ask 
Mostly interested in stability, ofc performance is also interesting.
Reason is that I’m going to build ESXi server(s) for 6 thin client users initially and perhaps grow to 10 later on.
Because the workload is single threaded based I will use threadripper instead of epyc for higher core clocks.
From what I’ve read the threadripper 1950X works fine today on esxi vsphere, was some issues in the beginning but they are fixed now.
Great work you have going on there Wendell, love and appreciate what you do.
