IO_uring Zerocopy Send Is Ready For Linux 5.20 Networking

[phoronix]

Phoronix: IO_uring Zerocopy Send Is Ready For Linux 5.20 Networking

The long in development work around IO_uring zero-copy send for the Linux kernel's networking subsystem is now ready for mainline...

https://www.phoronix.com/scan.php?page=news_item&px=Linux-5.20-IO_uring-ZC-Send

[S.Pam]

Io_uring continues to improve

[linuxgeex]

It'll be interesting to see how much this affects real-world use cases like Redis read from Optane.

[NobodyXu]

It'll be interesting to see how much this affects real-world use cases like Redis read from Optane.

This added feature here is zero-copy send for network socket, not for reading from Optane.

And yeah, it will be great if db like redis supports io-uring, but given that redis is often used as a in-memory database with persistent, I guess the performance improvements won't be as significant as something like postgresql, which mostly handles data larger than the ram.

[linuxgeex]

This added feature here is zero-copy send for network socket, not for reading from Optane.

And yeah, it will be great if db like redis supports io-uring, but given that redis is often used as a in-memory database with persistent, I guess the performance improvements won't be as significant as something like postgresql, which mostly handles data larger than the ram.

I didn't suggest that it was for Redis read from Optane. I used Redis read from Optane as an example of an application that was *fast enough* to see an impact due to improvements to the networking subsystem. ie DMA from Optane to memory, DMA from memory to the network device, marshalled by Redis, and comparing zero-copy to in-core copy. Now do you get it?

[NobodyXu]

I didn't suggest that it was for Redis read from Optane. I used Redis read from Optane as an example of an application that was *fast enough* to see an impact due to improvements to the networking subsystem. ie DMA from Optane to memory, DMA from memory to the network device, marshalled by Redis, and comparing zero-copy to in-core copy. Now do you get it?

Yep, sorry that I misunderstood you.

[Sadhu]

This added feature here is zero-copy send for network socket, not for reading from Optane.

And yeah, it will be great if db like redis supports io-uring, but given that redis is often used as a in-memory database with persistent, I guess the performance improvements won't be as significant as something like postgresql, which mostly handles data larger than the ram.

Dragonfly is api compatible alternative to redis and memcached using io-uring. They claim to be way faster. https://github.com/dragonflydb/dragonfly

[NobodyXu]

Dragonfly is api compatible alternative to redis and memcached using io-uring. They claim to be way faster. https://github.com/dragonflydb/dragonfly

Thanks for the information!
I didn't know it uses io-uring as its README.md says its performance improvements come from better data structure and multithreading.

[S.Pam]

Dragonfly is api compatible alternative to redis and memcached using io-uring. They claim to be way faster. https://github.com/dragonflydb/dragonfly

Quite nice haven't heard of it, but I will have to try!