The Brutal Performance Impact From Mitigating The LVI Vulnerability

[jrch2k8]

Well is official, i lived long enough to see my AMD FX 6100 beat the living crap of an core i9-9900ks

[Jumbotron]

"Thankfully the very latest Intel CPUs such as newer Coffee Lake parts, Comet Lake, and Cascade Lake are said to be largely unaffected while Ice Lake is not vulnerable at all."

Phoronix headline 1 year from today.....

"A new and even MORE malicious vulnerability was detailed today concerning Intel's latest generations of CPU's thus adding to the ever expanding list of Intel CPU's that are unfit for HPC, Hyperscale, and Cloud Based Computing. The parts effected by this new vulnerability, which Intel STILL denies is a problem, includes Coffee Lake, Comet Lake, Cascade Lake and Ice Lake which at one time was considered NOT VULNERABLE AT ALL."

[tuxd3v]

We start to live in strange times...
What is old, becomes new again.. this is what some would call "innovation"..

[Mario Junior]

mitigations=off and fuck this shit!
Also, if you need do this on Window: https://support.microsoft.com/en-us/...ulnerabilities

[yreew]

mitigations=off and fuck this shit!
Also, if you need do this on Window: https://support.microsoft.com/en-us/...ulnerabilities

since the mitigation is in the compiler, i dont think that works on this one, no?

[milkylainen]

Yeah. Pretty much expected it. From whatever I said in the previous LVI article.
The idea however is not to apply this "liberally", as in recompiling an entire program with mitigation.
The idea is to fix susceptible gadgets. But therein lies the problem. How to successfully track all susceptible gadgets?
Also, those gadgets need to be separate objects? Or can these mitigations be pushed with pragmas or attributes?
Either way, serializing critical instructions like this (esp. load) was obviously going to cause horrific performance numbers.
Mitigating like this is totally unusable in the real world.

[Toggleton]

I don't normally make these kinds of suggestions, but the results of random quad core x86_64 Intel processors from the past 20 years would be interesting.

Would be funny if core2quads started getting wins again.

https://openbenchmarking.org/result/2003133-VE-2003126NI94&obr_sor=y&obr_hgv=Core2Quad-Q6600
Intel Core 2 Quad Q6600(Launch Date Q1'07) 2.4GHZ

WINS:
No LVI Mitigation: 29 [61.7%]
LFENCE Before Indirect Branch: 10 [21.3%]
LFENCE Before RET: 4 [8.5%]
Core2Quad-Q6600: 2 [4.3%]
LFENCE After Load: 1 [2.1%]
LFENCE Load + Indirect Branch + Ret: 1 [2.1%]

LOSSES:
LFENCE Load + Indirect Branch + Ret: 19 [40.4%]
Core2Quad-Q6600: 17 [36.2%]
LFENCE After Load: 10 [21.3%]
LFENCE Before Indirect Branch: 1 [2.1%]
TESTS COUNTED: 47
----------------------
Harmonic Mean Of MiB/s Test Results
Harmonic Mean
MiB/s > Higher Is Better
LFENCE After Load ................... 21.11
LFENCE Load + Indirect Branch + Ret . 21.12
Core2Quad-Q6600 ..................... 65.90
LFENCE Before RET ................... 218.83
No LVI Mitigation ................... 219.91
LFENCE Before Indirect Branch ....... 220.12

Geometric Mean Of All Test Results
Result Composite
Geometric Mean > Higher Is Better
LFENCE Load + Indirect Branch + Ret . 186.87
LFENCE After Load ................... 190.03
Core2Quad-Q6600 ..................... 199.26
LFENCE Before RET ................... 739.72
LFENCE Before Indirect Branch ....... 763.88
No LVI Mitigation ................... 771.00

[TemplarGR]

Just disable hyperthreading (SMT) people. It is a garbage technology anyway. For most use cases it provides tiny benefits, especially for most desktop users, any benefits aren't worth the security holes. Better than using costly mitigations....

[TemplarGR]

Well is official, i lived long enough to see my AMD FX 6100 beat the living crap of an core i9-9900ks

Bulldozer has always been an elegant design. I am frustrated that AMD had to scrap it and go copy Intel's SMT design. Now Zen is just a me-too Core architecture with slightly worse IPC but more cores per dollar.

But hey, people cared about their "single threaded performance" AKA "muh garbage single threaded vidya gamez". Oh, and superpi, that's terribly important apparently.

I hope now that AMD begins enjoying a dominant position in the market, they can finally implement the plan of plenty tiny integer cores with plenty of large gpgpu cores, and HSA. It will be glorious.

[nokipaike]

just out of curiosity, the situation is theoretically dramatic in the same way even on windows right?