Use io_uring to batch handle clients pending writes to reduce SYSCALL count.

[lipzhu]

Description

This patch try to benefit the io_uring batching feature to reduce the SYSCALL count for Redis when handleClientsWithPendingWrites.
With this patch, we can observe more than 4% perf gain for SET/GET, and didn't see an obvious performance regression.

NOTE

• Since io_uring was adopted from kernel 5.1, if kernel doesn't support io_uring yet, it will use the origin implementation.
• This patch introduce the liburing dependency, it is installed in my local env, to keep it simple, I didn't include liburing dependency in this patch, the CI build may failed.

Test Result

Test Env

• OS: CentOS Stream 8
• Kernel: 6.2.0
• Platform: Intel Xeon Platinum 8380
• Base: 5b9fc46
• Server and Client in same socket.

Test Steps

1. Start redis-server taskset -c 0-3 ~/src/redis-server /tmp/redis_1.conf with below config.

port 9001
bind * -::*
daemonize yes
protected-mode no
save ""

2. Start redis-becnmark taskset -c 16-19 ~/src/redis-benchmark -p 9001 -t set -d 100 -r 1000000 -n 5000000 -c 50 --threads 4 to ensure redis-server CPU utilized is 1(fully utilized).

Test Result

With this patch, the QPS can increase 4.3%.
And both the cycles for each query and SYSCALL count are reduced as expected.
perf stat -p `pidof redis-server` sleep 10
Table is the normalized cycles/ins for each query.

	Base	Opt	Opt/Base-1
cycles/query	14,677	14,079	-4.1%
instructions/query	19,541	19,350	-1.0%
IPC	1.33	1.37	3.2%
CPU utilized	1	1	0%

bcc/syscount -p `pidof redis-server` -d 10
In this patch, io_uring_enter is the replacement of write.

SYSCALL COUNT	Base	Opt
read	1,950,855	2,081,797
write	1,950,726	0
epoll_wait	42,248	45,485
io_uring_enter	0	45,484
Total	3,943,829	2,172,766

From above data, the SYSCALL is expensive, by reducing the SYSCALL related instructions, IPC improves.

[zuiderkwast]

io_uring is very cool, but I don't understand it well enough yet. I commented on some things related to conditional compilation.

I think we can mark it as draft, because there are problems in CI that needs to be solved.

[zuiderkwast]

I think this will always be true. USE_IO_URING is not set to "no" anywhere. Why not ifeq ($(USE_IO_URING),yes)?

[lipzhu]

You are right, my original intention is to turn on io_uring by default unless the user manually turns off io_uring by make USE_IO_URING=no. ): It works on Linux, but for other platform, we should disable io_uring default.

Sure, I will change to ifeq ($(USE_IO_URING),yes) as you suggested. Thanks.

[zuiderkwast]

We can't include this on platforms that don't have it.

[lipzhu]

Yes, you are right, maybe we can solve this by ifeq ($(USE_IO_URING),yes). And use USE_IO_URING to protect it.

[lipzhu]

@zuiderkwast Thanks for your comments.

io_uring is very cool, but I don't understand it well enough yet. I commented on some things related to conditional compilation.

Yes, io_uring is very cool, really excited to see io_uring can help Redis on this scenario and gain perf. The scenario is very straightforward, appreciate that the community can help on this.

I think we can mark it as draft, because there are problems in CI that needs to be solved.

For the CI build failure, it is caused by the liburing is not exist in CI environment, I am thinking how to add the liburing dependency for Redis: manually copy liburing to deps/ folder or refer to the method of ssl. What's your suggestion?

[zuiderkwast]

I am thinking how to add the liburing dependency for Redis: manually copy liburing to deps/ folder or refer to the method of ssl. What's your suggestion?

I think we can auto-detect it like we do for C11 atomics here: https://github.com/redis/redis/blob/7.2.4/src/Makefile#L46. It creates a small file and tries to compile it.

[lipzhu]

I am thinking how to add the liburing dependency for Redis: manually copy liburing to deps/ folder or refer to the method of ssl. What's your suggestion?

I think we can auto-detect it like we do for C11 atomics here: https://github.com/redis/redis/blob/7.2.4/src/Makefile#L46. It creates a small file and tries to compile it.

The approach still need user manually install liburing even kernel has io_uring support, and sometimes the build env is not total same with runtime env, can we include https://github.com/axboe/liburing/ in deps module like jemalloc if license compliance?

[zuiderkwast]

user manually install liburing even kernel has io_uring support

Ok, now I understand. I think we can include it under deps if it is small. What's the total size of it? The core team needs to make the final decision.

An idea is to introduce it in two steps. First, as a beta feature, we can we require the user to manually install liburing. We can modify one of the CI jobs to install and use it in tests. Later, when it's found to work well, it can be enabled by default and included under deps.

[lipzhu]

user manually install liburing even kernel has io_uring support

Ok, now I understand. I think we can include it under deps if it is small. What's the total size of it? The core team needs to make the final decision.

With liburing, the total size of redis-server binary changed from 21279160 to 21451704.

An idea is to introduce it in two steps. First, as a beta feature, we can we require the user to manually install liburing. We can modify one of the CI jobs to install and use it in tests. Later, when it's found to work well, it can be enabled by default and included under deps.

Agree, I just upload the liburing-2.5 to deps to check if the CI could success, this is a temporary solution. The final solution need core team make decision.

[lipzhu]

Update: The patch makes below 4 unit tests hang, the others passed, I am trying to fix it, if any help will be appreciated. :)

• Client output buffer hard limit is enforced
• Client output buffer soft limit is enforced if time is overreached
• Client output buffer soft limit is not enforced too early and is enforced when no traffic
• slave buffer are counted correctly

BTW, @madolson @oranagra @zuiderkwast @enjoy-binbin do you think this make sense for Redis to adopt io_uring from this scenario?

[CLAassistant]

All committers have signed the CLA.

[oranagra]

@lipzhu can you save me some time and improve the top comment describing the change in the flow.
i.e. i imagine it adds the buffers to be written without a syscall (and without copying the data), and then later waits for the writes in one system call before proceeding to the event loop? please add some details to explain the flow before i go read the code.

additionally, i see the benchmark mentions the reducing of cycles per commands, but not any improvement in throughput or latency, can we add one? how will it be affected by adding a pipeline (-P) to the benchmark?