Introduce io_uring support to ClickHouse

[chhetripradeep]

Linux kernel has released a new asynchronous I/O APIs : io_uring and has added support for all network and disk I/O syscalls over last 1.5 year and it looks very promising.

Few of the other project has started adopting it and have seen performance benefits:

• Ceph: os/bluestore: introduce new io_uring IO engine ceph/ceph#27392
• Qemu: https://lists.gnu.org/archive/html/qemu-devel/2019-07/msg06733.html

Many others frameworks like Netty are planning to adopt it too: https://twitter.com/normanmaurer/status/1257597649705893888

In context with the scale at which everyone runs clickhouse, I think it will be of great performance improvement.

[bobrik]

Looking at flamegraphs for some queries I see quite a bit of copy_user_enhanced_fast_string, which is supposed to be alleviated with io_uring, as far as I understand:

[alexey-milovidov]

@bobrik copy_user is also eliminated with:

• Linux AIO that is enabled with min_bytes_to_use_direct_io setting;
• mmap that is enabled with min_bytes_to_use_mmap_io setting.

[alexey-milovidov]

Now we have the notion of read_method to select between various read engines.
And we can start experimenting with io_uring.

CC @azat maybe you will find this interesting?

[azat]

CC @azat maybe you will find this interesting?

Yep, this idea already visited me (added these to my internal to-do/to-look list)

[alexey-milovidov]

There was an experiment adding uring by @sauliusvl
Unfortunately, it proves to be unsustainable.

There is only a marginal improvement in performance. But the code becomes way more complicated. It became so complicated that even an experienced C++ engineer (the author of the code) cannot figure out why there are rare hangs of queries (found by our automated testing before the release).