Docker Performance With KPTI Page Table Isolation Patches

[phoronix]

Phoronix: Docker Performance With KPTI Page Table Isolation Patches

Overall most of our benchmarks this week of the new Linux Kernel Page Table Isolation (KPTI) patches coming as a result of the Meltdown vulnerability have showed minimal impact overall on system performance. The exceptions have obviously been with workloads having high kernel interactions like demanding I/O cases and in terms of real-world impact, databases. But when testing VMs there's been some minor impact more broadly than bare metal testing and also Wine performance has been impacted. The latest having been benchmarked is seeing if the Docker performance has been impacted by the KPTI patches to see if it's any significant impact since overall the patched system overhead certainly isn't anything close to how it was initially hyped by some other media outlets.

http://www.phoronix.com/vr.php?view=25799

[s_j_newbury]

I don't think it really has been over-hyped, perhaps misunderstood or misinterpreted. That it hasn't devestated micro benchmarks across the board isn't surprising nor what you would expect to see from the additional overhead incurred by KPTI. Where I suspect it will be felt is with systems which are designed to run at high load, rather than number crunching, or relatively simple single process tasks.

[linuxgeex]

Michael Yes it looks like real-world performance of applications is largely unaffected.

I found this interesting - Chris Ramsayer (THG) found that Optane 4kQD16 writes dropped from 395k to 71k IOPS in the worst case scenario:

http://www.tomshardware.com/news/mic...nce,36236.html

"Synthetic workloads show what could be possible--in this instance, the worst-case scenario [...] When it comes to modern storage, the software is the bottleneck rather than the hardware components."


But it's only certain cloud infrastructure and business logic that's taking a hit, because those are the only apps that are hitting such high syscall rates.

Memcached must be hit even harder than these Optane results since it's an in-memory database. I'm sure other in-memory DB apps like SAP Hana would also take huge hits but those apps are so specialized that they will be run on dedicated hardware and can be run with KTPI disabled, whereas Memcached is usually run on shared public-facing servers.

I wonder how AWS, Azure, AppEngine, Bluemix, and Twitter are coping with the impact.

[Leopard]

Meltdown patches affected performance of Epic Games's Fortnite servers.

https://www.epicgames.com/fortnite/forums/news/announcements/132642-epic-services-stability-update

See the spike at cpu usage.

[andyprough]

Over hyped end-of-world scare headlines by most of the media.

[vegabook]

Over hyped end-of-world scare headlines by most of the media.

"Idiotic head buried in sand wishful thinking comment failing to acknowledge the obvious reality that our shared-VM infrastructure-style workloads have just been comprehensively trashed"

https://www.epicgames.com/fortnite/f...ability-update

[andyprough]

"Idiotic head buried in sand wishful thinking comment failing to acknowledge the obvious reality that our shared-VM infrastructure-style workloads have just been comprehensively trashed"

https://www.epicgames.com/fortnite/f...ability-update

Oh noes. We will lose 1-2% performance. Same amount we typically lose with a bog-standard kernel regression. The world is ending. Run for your lives.

The whole paradigm for speculatively caching data has been screwed for decades. About time something happened to force a re-think of the entire concept.

[ramrod]

Oh noes. We will lose 1-2% performance.

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

[gururise]

Machine Learning (especially Scikit-Learn) performance can suffer up to 40% performance drop:

https://medium.com/implodinggradients/meltdown-c24a9d5e254e

[gururise]

Also the Meltdown mitigation patch caused 100% load increase on EPIC Games' Fortnite's servers and service degradation:
https://www.epicgames.com/fortnite/f...ability-update