2990WX -- 4.1GHz --24/7 -- All-Cores -- Shared Benchmark Script -- UPDATE-9-25-18 --Kernel-4.19.0-rc5 -- What 'make FLAGS' do you all usually use when compiling the Linux kernel and why?

I didn’t try 2800, but I did run at 2933, 3000 and 3200 with 64 and 2933/3000 with 64. Seemed like a pretty linear improvement at each step. More IF speed = more goodness.

One thing I’ll say from Intel memory OC world, DDR4 re-tries can silently eat into your memory performance when you’ve pushed it to the ragged edge. Just before you start getting reported errors, you are getting silent re-tries from CRC errors at the DDR4 PHY level.

This is often what’s going on when you start seeing performance degradation at higher speeds without reported errors (assuming your net latency is reduced by the higher speed and not just trading clock rate for cycles on primary/secondary timing).

I made your suggested changes to the benchmark and tested. Using the $TEST_ROOT was about 4 seconds slower and doubling the threads was about 10 seconds slower. However, I do agree it could benefit from a few more threads so I am adding 1 at a time to see where the sweet spot is. I do encourage you to download the benchmark and tweak it to see if you are able to further optimize it. Your hardware may respond differently than mine. Thank you for taking the time to help.

I was able to shave off 3 more second by adding 8 more threads thanks to your suggestion.

Yes. I agree. When I have time I will do more testing to find the best memory speed and timings.

you could shave off more time by running it without x.

Yes definitely.

Hi Wendell,

Would you please share with me the .config that was used to compile the kernel in 30 seconds?

haven’t gotten to this yet, and I wouldn’t know where the phoronix test suite keeps it without some digging

Is it something other than what you get with make defconfig? I’d be surprised if phoronix wanted to get more complicated than that. They have enough to deal with…

You’ll find .config in the root of the kernel build tree after that.

That runs as follows on my 7980xe (4.5GHz - raid0 nvme 960 pro for reasons):
git clone…
cd linux-stable
make defconfig
time make -j36

Kernel: arch/x86/boot/bzImage is ready (#1)

real 0m30.283s
user 11m12.731s
sys 1m28.189s

So, I’d expect the 2990wx produces similar results on this getting us to the ~30s mark.

1 Like

I would be interested to see how quickly the 7980xe could run the benchmark that I shared to put things in perspective.

<<---------------------------------Results-------------------------------->>

<<------------------------------------------------------------------------>>
Sat Sep 22 16:43:27 EDT 2018
Compiled kernel ---> 00:00:03:35
<<------------------------------------------------------------------------>>

Toasty 80C max package temp (delid - lm - custom water - quiet fan profile though) :wink:

kernel builds do keep it busy:
kernel_cpufreq

(disclaimer: not intended to be flame-war bait… this chip is delidded, hitting 80C, liquid metal, drastically OC’d from its stock settings, etc… This is how I’ve run it 24/7 for about a year now, but hopefully people can keep things in perspective.)

1 Like

Very nice! Thank you so much for taking the time to provide me something to compare my results to.

I modified my cpu info script. Check it out. Let me know if you see any improvements I could make.
!!START_CODE!!
while [[ 1 ]]; do clear && echo -e “Threadripper 2990WX – Cores: 32 Threads: 64 – GET SOME” && awk ‘/processor/ || /cpu\ MHz/ { print “Proc:”$3":"$4 }’ /proc/cpuinfo | sed -r ‘s/cpu//g’ | paste - - | sed -r ‘s/Proc:://g’ | sed ‘s/’\t’//g’ | awk -F. ‘{ print $1 }’ | sed -r ‘s/::/–>/g’ | paste - - - - - - - - && sleep .1; done
!!END_CODE!!