Question Zen 6 Speculation Thread

[Io Magnesso]

The sad part is that, what makes computer systems feel more and more anemic as time goes on isn't that the OS itself is slowing down, it's all the constant security vulnerability mitigations that keep getting piled on. The computers that felt super quick on early Windows 10 at work a few years ago are just crawling these days. You crack open performance monitor and dig around and you see all sorts of auditing activities, antivirus scanning, memory randomization efforts, intentional cache dirtying, memory address scrubs, access profiling etc.

Systems don't really idle that often. There's always something in the background with something to do.

That's right, early Windows 10 worked fine even with 4GB of RAM.

 

[Markfw]

AMD hardly had any "fixes" to slow it down, I don't know where this "slow computer" comes from.

 

[Io Magnesso]

AMD hardly had any "fixes" to slow it down, I don't know where this "slow computer" comes from.

What does that mean?

 

[LightningZ71]

A lot of it isn't processor specific. OSes in general are adopting more security fundamentals to be resistant to zero day hardware exploits. Newer languages now do memory access obfuscation and randomization by default in the background. It's a fundamental part of RUST for example.

All of that eats up processor cycles and memory access cycles.

 

[MS_AT]

memory access obfuscation and randomization by default in the background. It's a fundamental part of RUST for example.

Could you link to RUST docu, that supports these claims? I mean its first time I see it, that they have built memory obsfucation and randomization into the language.

Things like ASLR (address space layout randomization) is language agnostic afaik and depends on the OS defaults, but you can overrule them.

 

[igor_kavinski]

Safety comes at a price.

https://dl.acm.org/doi/fullHtml/10.1145/3551349.3559494

 

[MS_AT]

View attachment 126378

Safety comes at a price.

https://dl.acm.org/doi/fullHtml/10.1145/3551349.3559494

I never thought I would defend rust but to be fair, quouting from the article

"Names of the algorithms can be found in Figure 1 and all the programs are contributed by irrelevant third parties."

So they don't provide the sources of the algorithms they compare. They claim they verified them themselves but it makes it impossible to reproduce nor to tell if they were written in idomatic ways that make them easy for compiler to optimize. I mean we have to trust the authors were familiar with the best way to write in each language.

https://benchmarksgame-team.pages.debian.net/benchmarksgame/fastest/rust.html this shows more modern take if I understand correctly. At least you can see the code. The results are more in line with each other.

Also comparison should be done between clang compiled c and rust, since both are build on top of llvm, so it is better way to single out language specific influence

 

[Markfw]

A lot of it isn't processor specific. OSes in general are adopting more security fundamentals to be resistant to zero day hardware exploits. Newer languages now do memory access obfuscation and randomization by default in the background. It's a fundamental part of RUST for example.

All of that eats up processor cycles and memory access cycles.

The vast majority is CPU wise, and Intel has way more. For the OS, they also check to see what the CPU is to decide if a certain fix is required. So you have any benchmarks to show anything concrete ?

 

[LightningZ71]

I didn't do full benchmark workups at the time because it wasn't in the scope of my investigation. We were trying to nail down where the slowdowns were coming from to make sure it wasn't intrinsic to any of our installation process program suite. Both AMD and Intel systems showed similar degrees of performance deterioration for user functions between a historic image and our current one with the same software installed. I will grant that this was last looked at before MS did their famous AMD mitigations patch a while back where they straight up admitted to enforcing unneeded mitigations on certain AMD processors, but in a cursory look, that didn't make a big change for our numbers.

In general, both Windows and Linux are having to do extra work to improve system security. IMHO, Linux distro seem to age a bit more gracefully because that's what I see in production, but it's not entirely free of this issue. And, no, I can't just choose to turn off all mitigations. That would worsen our organization's security posture. Also, it's not ALL on the OS (excluding microcode changes) as the software itself has to do more work verifying inputs don't overrun buffers and verifying their own internal file integrity. Unless they are patching out performance bugs, security updates often worsen performance.

It also depends on the processor on how well it comes with OS changes. We had a large batch of 7th Gen i5 processors in production that we're getting absolutely hammered by this. Stuff that had limited L3 amounts and less than 8 threads was really suffering. Protecting memory access causes tons of process stalls. 4 core HT processors made a big difference. It was less noticeable on the Ice Lake and tiger lake laptops that we looked at, even with their I/o mitigations in place.

 

[igor_kavinski]

It was less noticeable on the Ice Lake and tiger lake laptops that we looked at, even with their I/o mitigations in place.

Ice Lake is the context switch latency champ.

ctx_clock Benchmark - OpenBenchmarking.org

openbenchmarking.org

Random 4K I/O involves context switches. Doesn't explain better performance of Tiger Lake results you saw.

 

[LightningZ71]

As compared to 7th Gen I5? The Tiger Lake processors had more L3 and HT on top of an improved core. Much more resilient to memory access stalls.