Linux Achieves 5.1M IOPS Per-Core With AMD Zen 3 + Intel Optane

[phoronix]

Phoronix: Linux Achieves 5.1M IOPS Per-Core With AMD Zen 3 + Intel Optane

Linux kernel developers have been working tirelessly to squeeze more performance out of IO_uring and the block / I/O code in general. IO_uring lead developer Jens Axboe who also serves as the Linux block subsystem's maintainer (among other roles and major contributions over the years) has used his system as a baseline for evaluating such kernel improvements. He's now moved to using AMD Zen 3 while sticking to Intel Optane storage and is seeing a mighty speed boost out of AMD's latest processors...

https://www.phoronix.com/scan.php?page=news_item&px=Linux-5.1M-IOPS-Per-Core

[Danny3]

Incredible, these optimizations are really amazing !
I just wish more software would use this IO_uring.

[ermo]

When both Jens Axboe and Linus Torvalds use prosumer-level AMD hardware (with ECC RAM) for their Linux work, you know AMD must be doing something right...

EDIT: And Greg Kroah-Hartman (hat tip to Teggs)

[onlyLinuxLuvUBack]

intel was only 3 million ?

so
AMD zen3+ > intel ?

[perpetually high]

When both Jens Axboe and Linus Torvalds use prosumer-level AMD hardware (with ECC RAM) for their Linux work, you know AMD must be doing something right...

My next deskop will absolutely have ECC memory. Glad Linus made a point about that.

[bug77]

All signs seem to indicate with Alder Lake he'll be able to push past 5.1M.

[ermo]

intel was only 3 million ?

so
AMD zen3+ > intel ?

He was on Zen 2 (Threadripper 3970X) before according to the article.

So it's really Zen 3 > Zen 2 for Jens Axboe's use case.

Linus has made some choice remarks on intel's product segmentation choices and pricing structure. It basically boils down to AMD having a pricing structure (due to their Zen chiplet approach) where if you pay twice as much for a CPU, you generally get twice as much performance (in the form of double the number of cores at the same or higher speeds).

When Linus made his comments, he also stated something along the lines that if you want ECC support on Intel (= Xeon) you pay five times as much for twice the performance, which means that the net effect is that consumers are generally encouraged to buy non-ECC memory if they want Intel systems at prices that are competitive with equivalently priced AMD systems (which also support ECC, even if only "unofficially").

His posts in that thread are worth a read IMHO.

(source).

[cl333r]

io_uring can't even do 800 million IOPS per core, what a pathetic loser.

[onlyLinuxLuvUBack]

All signs seem to indicate with Alder Lake he'll be able to push past 5.1M.

Where can you sell me this drive ? and does it come with a bridge ?

[sophisticles]

When both Jens Axboe and Linus Torvalds use prosumer-level AMD hardware (with ECC RAM) for their Linux work, you know AMD must be doing something right...

Or they have corporate sponsorship and they didn't buy these systems themselves.

I have no proof that they received their systems for free, and I'm not saying that AMD's high core count processors are not impressive but I am weary of the "appeal to authority" tone of this post, where the implication is if they are using it then it much be great.