Async Buffered Reads Support Yielding Promising Results

[phoronix]

Phoronix: Async Buffered Reads Support Yielding Promising Results

Linux I/O expert Jens Axboe who oversees the kernel's block layer and is employed by Facebook while working on IO_uring and other storage innovations has recently been working on async buffered reads support...

http://www.phoronix.com/scan.php?page=news_item&px=Linux-Async-Buffered-Reads

[S.Pam]

The reduced io load and cpu load is also something I've seen with samba and io_uring. Rather cool, especially on low end home servers

Just add the following to your smb.conf
[someshare]
vfs objects = io_uring

Note that you need samba 4.12.3. Previous versions had a data loss bug with some Windows clients.

Edit: Samba with io_uring works with most newer kernels. This patch set for 5.8 is to bring the benefits to all async io

[Steffo]

Cool! Which kernel version will contain this patch set?

[Michael]

Cool! Which kernel version will contain this patch set?

Linux 5.8 as of a few minutes ago now queued in linux-block.

[CochainComplex]

Benchmarks ?

[MastaG]

5.8 Will be great !

[bug77]

5.8 Will be great !

Show me a version of the Linux kernel that is not great

[sindr]

Show me a version of the Linux kernel that is not great

All Linux kernels are great and terrible things.

[boxie]

Show me a version of the Linux kernel that is not great

Every kernel version ending in .0 ? :P

*runs away in a zigzag fashion, laughing maniacally*

[NotMine999]

Benchmarks ?

Keep looking.

The only performance testing stats I saw were in the LKML post by Jens Axboe. Just stats; no testing methodology, no listing of how many test repetitions were performed to obtain those results, no pointer to the testing app claimed in the LKML. From a scientific method POV it was useless for critically and objectively evaluating performance.

Sadly, I don't call what Axboe posted in the LKML to be "benchmarks"; it seemed like more of a passing comment to support his wish to have that code merged.

preadv for comparison:
real 1m13.821s
user 0m0.558s
sys 0m11.125s
CPU ~13%

Mainline:
real 0m12.054s
user 0m0.111s
sys 0m5.659s
CPU ~32% + ~50% == ~82%

This patchset:
real 0m9.283s
user 0m0.147s
sys 0m4.619s
CPU ~52%