IO_uring Gets New Features & Speed-Ups With Linux 5.18

[phoronix]

Phoronix: IO_uring Gets New Features & Speed-Ups With Linux 5.18

The Linux 5.17 kernel will be released tomorrow and in turn that will kickoff the start of the Linux 5.18 merge window. Linux block subsystem maintainer Jens Axboe has already begun submitting his feature pull requests for this next kernel, including the IO_uring updates...

https://www.phoronix.com/scan.php?page=news_item&px=Linux-5.18-IO_uring

[S.Pam]

Seems no end to how much improvements we can get in io_uring.

I'd love to see more applications using it for network sockets as well as disk io. For example servers like Tor relays, Web servers and others with lots of connections would benefit well from this.

[Rexerex]

For example servers like Tor relays (...) would benefit well from this.

Tor probably should wait for some time until all bugs are found due to security reasons.

[JEBjames]

Michael

Articles could be optimized a further 2-10% by removing all spaces.

Typo "improvementswith".

[kozman]

Seems no end to how much improvements we can get in io_uring.
I'd love to see more applications using it for network sockets as well as disk io. For example servers like Tor relays, Web servers and others with lots of connections would benefit well from this.

Just think of how much faster this will make servers be able to use AI, algorithms and serve ads to us. Joy.

[coder]

Seems no end to how much improvements we can get in io_uring.

A slight apprehension I have is that specific io_uring code paths and usage models are getting optimized potentially at the expense of others.

[SvenK]

Just think of how much faster this will make servers be able to use AI, algorithms and serve ads to us. Joy.

Well we need direct access from GPUs to Data on HDDs or SSDs.

[coder]

Well we need direct access from GPUs to Data on HDDs or SSDs.

Nvidia has some proprietary way of doing that, called GPUDirect. I don't know how much potential there is for a general solution.

And it's not really HDDs that are a use-case. The main one is fast NVMe SSDs (or RAIDs of them).

[dragorth]

A slight apprehension I have is that specific io_uring code paths and usage models are getting optimized potentially at the expense of others.

This is my concern as well. Hopefully there are regression tests for current systems, but I don't know.

[S.Pam]

Hopefully there are regression tests for current systems

It's called Phoronix!