io_uring performance 40% better than kqueue and epoll !

Status
Not open for further replies.
zoujiaqing

zoujiaqing

Io_uring: New Linux IO interface (I think it like IOCP)

Linux 5.1 introduces a new high-performance interface called io_uring. It’s easy to use and hard to misuse user/application interface. Io_uring has an efficient buffered asynchronous I/O support, the ability to do I/O without even performing a system call via polled I/O, and other efficiency enhancements. This will help deliver fast and efficient I/O for Linux.\

Io_uring permits safe signal delivery in the presence of PID reuse which will improve power management without affecting power consumption. Liburing is used as the user-space library which will make the usage simpler. Axboe’s FIO benchmark has also been adapted already to support io_uring.

hub.packtpub.com

Linux 5.1 out with Io_uring IO interface, persistent memory, new patching improvements and more! | Packt Hub

Linus Torvalds, announced the new Linux 5.1 version. It has new features like Io_uring and improvements to storage and security
hub.packtpub.com hub.packtpub.com

liburing - io_uring library

git.kernel.dk git.kernel.dk
 
That's a bit like saying that a 16 Gb memory stick is faster than a 4-core CPU.
They both revolve around the sun with the same velocity.

If you believe that FreeBSD needs an io_uring analog, you can propose that on the freebsd-hackers mailing list.
 
Please explain how this is relevant, how the measurements were done, and what you intend to accomplish by writing this.
 
ralphbsz said:
Please explain how this is relevant, how the measurements were done, and what you intend to accomplish by writing this.
Click to expand...

This.:eek:
It smells like a troll, but it might just be language barriers. Still it baffles me what a linux in-kernel hack has to do with freebsd?
 
ralphbsz said:
Please explain how this is relevant, how the measurements were done, and what you intend to accomplish by writing this.
Click to expand...
If you mean my post - detailed description of testing benchmarking Debian from 7 (3.2 kernel announced 5 may 2013) to 11 (current 5.x kernel) - here and includes and IO tests on page 2.
Margin in the random reading and writing doesn't equal + or - 40% from absolute values.
I don't see IO software revolution for increasing performance IO from Linux 3.2 -5.x kernel in these tests.
 
Vadim_Mkk said:
If you mean my post - detailed description of testing benchmarking Debian from 7 (3.2 kernel announced 5 may 2013) to 11 (current 5.x kernel) - here and includes and IO tests on page 2.
Margin in the random reading and writing doesn't equal + or - 40% from absolute values.
I don't see IO software revolution for increasing performance IO from Linux 3.2 -5.x kernel in these tests.
Click to expand...

The benchmarks on phoronix are dubious at best when translated to real world and are Linux - centred (of course, as that's what phoronix exists for, to pander to the Linux fanboy). Any comparison by phoronix between Linux and FreeBSD should be taken as bogus.

Most benchmarks in general are a waste of your time unless it's your system and servers you're working on.
 
The Phoronix tests are absolutely useless. There is no thorough analysis of what the bottleneck was, and what we're actually measuring. Actually, they're worse than useless: people might believe the results as relevant, and organize their lives around it.

I agree that there is a problem in today's OS stack: we now have devices that are extremely fast for small IOs (around microsecond), and the overhead of the traditional stack (from userspace pread/pwrite all the way through system and device driver) is too heavyweight. But this is only relevant for a tiny fraction of all applications, in particular the backend implementation of databases and file systems. About 99.99% of all computer users and programmers should not even think about this stuff; they're more likely to break something (for example themselves).
 
I think I am seeing a trend with the OPs topics:
  • What's state Gnome with Wayland on FreeBSD?
  • rpm4 was support for building rpm without Berkeley DB (–disable-bdb)
  • Do you have tools like fedora's flatpack?
  • How to use plasma 5 on wayland?
  • etc...
So it sounds like either they are pining for Linux or they are simply on the wrong forums.
 
Status
Not open for further replies.
Back
Top