@dreamlike-ocean let me know once you did run some benchmarks

Would be great while benchmarking, to collect a statistics of the different cases eg total in memory hit/sec, syscalls enter due to pending cqes or sqes to submit and eventually flame-graphs to make sure we have enough concurrent and small I/O

Any news @dreamlike-ocean ?

Any news @dreamlike-ocean ?

Sorry, I’m currently on vacation. I’ll move this forward once I’m back.

@dreamlike-ocean enjoy your time off!

Benchmark Results

Test Setup

• Simple HTTP/1.1 server returning a 512-byte response body
• Using buffer ring + zero-copy write
• Load testing tool: oha

1. High Concurrency Load Test (oha -c 500 -z 30s)

strace statistics for io_uring_enter:

Conclusion: No performance regression under high concurrency, but the optimization shows minimal effect. This is expected — under saturated IO, needSubmit() almost always returns true because the SQ continuously has pending SQEs (count > 0), or IORING_SQ_TASKRUN is persistently set due to IORING_SETUP_DEFER_TASKRUN keeping deferred task work pending in the kernel.

2. Non-blocking Path Validation (IO idle + busy task queue)

To isolate and validate the effect of needSubmit() on the non-blocking path, I designed a scenario that forces the event loop to always take the non-blocking path (submitAndClearNow) with zero IO traffic. This is achieved by submitting an infinitely recursive task to each worker EventLoop, keeping hasTasks() permanently true and canBlock() permanently false:

 worker.forEach(Main::executorInfiniteTask);

private static void executorInfiniteTask(EventExecutor executor) {
    executor.execute(() -> executor.execute(() -> executorInfiniteTask(executor)));
}

strace statistics for io_uring_enter (30 seconds):

The remaining 65 calls in PATCH come from the acceptor thread (which does not have the infinite task) performing normal blocking waits. The OLD version's low3.4ms per call confirms these are empty submits (submit=0, minComplete=0) returning quickly from the kernel — essentially wasted syscalls. The PATCH version's 18.6ms per call confirms these are exclusively legitimate blocking waits from the acceptor.

Conclusion: On the non-blocking path with no IO activity, needSubmit() effectively eliminates unnecessary io_uring_enter syscalls, achieving an86.6% reduction.

raw benchmark result
benchmark.zip

@franz1981

Looks great 😃 today will take a look more deeply

@dreamlike-ocean thanks a lot!

Auto-port PR for 5.0: #16397