Docker Performance With KPTI Page Table Isolation Patches

[PuiuCS]

There are a lot of reports of HVMs and Redis nodes running on AWS having big CPU usage spikes (more than 30%).

Besides Epic games having login server issues, it seems other companies are having the same problems. For example CCP posted the CPU usage for their ESI production servers (their new API server) and after the patch they were getting as high as 100% more CPU usage.

[linuxgeex]

Machine Learning (especially Scikit-Learn) performance can suffer up to 40% performance drop:
https://medium.com/implodinggradient...n-c24a9d5e254e

then they'll need to switch from a syscall-heavy model to a shared memory model. If they're relying on network throughput then they'll need to implement ATM/epoll. If they're bottlenecked entirely by latency across nodes then they're in trouble.

[nanonyme]

thank you. this is an information that I would expect to find in phoronix benchmark. not this usual compile time related masturbation.

I've been reading people getting up to 30% degradation with Redis. Michael, can you do benchmarks with various database systems with the patch? Probably in AMI-like setup because that's what most people who cloudified their internal systems are running

[pete910]

From what I've been reading the drops have been when the relevant board bios has been updated with their fixes thats needed where the large drop in perf has occurred .

Admittedly this was on windows based systems.

[linuxgeex]

I've been reading people getting up to 30% degradation with Redis. Michael, can you do benchmarks with various database systems with the patch? Probably in AMI-like setup because that's what most people who cloudified their internal systems are running

There's been multiple Redis tests already. Go read.

[ChamPro]

Thanks Michael. This Docker benchmark was exactly the one I was looking for pertaining to Meltdown.

[gururise]

There is supposedly some microcode fixes coming from Intel. I wonder if the KPTI kernel fixes + the Intel microcode fixes cause even worse performance degradation.

TechSpot is reporting some significant disk slowdowns and further gaming slowdowns on Windows when using the KPTI patch along with the new Intel microcode that makes the branch predictor less aggressive: https://www.techspot.com/article/155...mance-windows/

Would be interesting to see the impact on Linux with the combination of KPTI kernel patch + Intel microcode.

[andyprough]

I'm guessing you didn't look at what he linked...

Oh noes. My online game server login isn't working. The world is ending.

[vegabook]

I think one of the main issues is not the _average_ workload hit, but the standard deviation. From the benchmarks I have seen so far, many workloads are hit by less than 5%, and then all of a sudden, some by 30% or more. This is the killer problem. The uncertainty. In many workloads you have to plan for the 99% percentile worst case and even if the average is a small hit, we're seeing some very big hits on the outliers.