Achieving higher throughput

[bits0rcerer]

From time to time I am playing around with iouring for networking.. espacially tcp connections.
Today I did a basic comparison on good ol` tokio vs iouring (using registered buffers, buf_ring and recv_multishot). My goal was to receive as many bytes as possible per second.

I was pretty sure I wont reach tokio`s battle-tested performance in an afternoon but even so I was surprised that tokio reached ~2 times the throughput.

Tokio: ~110 Gbit/s
io-uring: ~55 GBit/s

Network: loopback (127.0.0.1)
Kernel: 6.9.6-zen1-1-zen
CPU: AMD Ryzen 9 5900X (24) @ 3.700GHz

I used btop to meassure the throughput and #290 for the iouring buf_ring.

Maybe someone here can point out any obvious skill issues on my side or give some general advice.

iouring_tcp_sink.rs

tokio_tcp_sink.rs

tokio_tcp_src.rs

[DXist]

Your io_uring based program is single threaded while for tokio you've enabled multi-threaded lifetime.

What you can do:

• organize shared nothing approach for request processing:
  • one worker & io_uring per core
  • listen sockets are opened with SO_REUSE_PORT option, kernel does connection load balancing
  • use fixed file descriptors to remove per IO reference counting overhead
  • use fixed buffers to remove per IO memory pinning overhead
  • trade energy efficiency for throughput by using io_uring submission queue polling
  • configure network card, irq balancer and task scheduling to collocate packet routing, IRQ/network stack handling and application processing on the same CPU core to minimize cross CPU traffic.

[bits0rcerer]

Thank you for taking a look :)

• one worker & io_uring per core

I thought I already did exactly that by creating an io_uring in every thread (also with single issuer).

• use fixed buffers to remove per IO memory pinning overhead

I thought I already did that with buf_ring.

• listen sockets are opened with SO_REUSE_PORT option, kernel does connection load balancing

I'll definitely try that!

• use fixed file descriptors to remove per IO reference counting overhead

Same here

• trade energy efficiency for throughput by using io_uring submission queue polling

Same here as well

• configure network card, irq balancer and task scheduling to collocate packet routing, IRQ/network stack handling and application processing on the same CPU core to minimize cross CPU traffic.

Do you know if there are specific settings that play well with io_uring or is that general advice?

[DXist]

I like this series of articles:

• receiving side:
  https://blog.packagecloud.io/monitoring-tuning-linux-networking-stack-receiving-data/
  illustrating diagrams - https://blog.packagecloud.io/illustrated-guide-monitoring-tuning-linux-networking-stack-receiving-data/

• sending side - https://blog.packagecloud.io/monitoring-tuning-linux-networking-stack-sending-data/

The first article about receiving side describes multi queue NIC configuration for Receive Side Scaling.
An option to consider is to load balance connection flows using NIC and reserve a separate NIC queue + listening socket in each thread. This setup is similar to ScyllaDB.