The Performance Impact Of AMD Changing Their Retpoline Method For Spectre V2

[phoronix]

Phoronix: The Performance Impact Of AMD Changing Their Retpoline Method For Spectre V2

Made public this week was the Spectre-BHB / BHI vulnerability and while only Intel and Arm processors are currently believed to be impacted, in the course of that research the folks at VUSec discovered AMD's current Retpoline strategy for Spectre V2 mitigations is not adequate. This has led to a change in behavior for AMD processors and is already applied to the Linux kernel. Here is a look at what it means for desktop and server performance due to the change in return trampoline handling.

https://www.phoronix.com/vr.php?view=30973

[waxhead]

Will mitigations=off give and benefit what so ever if the CPU is patched with microcode?!

It would be interesting to see if kernel mitigations are faster or slower than microcode fixes, but then you would need a CPU without the microcode update.

[birdie]

I don't understand how to read these graphs.

Are we (AMD users) now slower, faster, roughly the same as we were before?

[numacross]

I don't understand how to read these graphs.

Are we (AMD users) now slower, faster, roughly the same as we were before?

It depends on the CPU and workload. I am assuming you mean the left-right graphs. On the first graph the new code is slightly faster in browsers while a lot slower in context switching. In the second everything is slightly slower with the new code and the last one is a mixed bag.

[rclark]

I too would like to know what performance difference for " mitigations=off" in comparison to all the so called security patches..... All my desktops and laptops are running AMD Ryzen processors.

[Danny3]

I don't understand how to read these graphs.

I find them confusing too!

[eoerl]

for what it's worth I find them very good
- It's meant to show all the tests at once
- left or right means left is faster, or right is faster, with the corresponding percentage
- it's sorted from the most affected (biggest balance) to the less affected

[KaoDome]

You're not alone birdie and @Danny3!!

What the hell is up with those graphs? I mean, what are they tracking? Worse is the further away it is from the center line? Because I see +X% at both sides of it. I'm a simple gal, box and whisker plots are the most I can understand at first glance haha

[jochendemuth]

On the 8-core AMD Ryzen 9 5900HX the test with the biggest negative impact was the generic Context Switching test of the Stress-NG suite with the prior default being 22% faster.
On the 16-core AMD Ryzen 9 5950X the test with the biggest negative impact was the generic Context Switching test of the Stress-NG suite with the prior default being 54% faster. That's more than double the impact on twice the amount of cores. I theorize that context switching on higher cores are heavily impacted by the patch.

I would love to see the result for this test on the 8-core EPYC (actually, I really want to see the impact on the 64-core EPYC, but that config wasn't tested), but the result is not listed.

[8r34k0u7_57y13]

I theorize that context switching on higher cores are heavily impacted by the patch.

I'll add to that, that I think it's going to be specifically worse across dies and possibly even worse across sockets. It'd be really interesting to see how Zen2 handles it, with 4c CCX - or even worse a 3c CCX part like 3600/3900/3960X