27us roundtrip is not really state of the art for zero copy IPC, about 1us would be. What is causing this overhead?
What exactly does that roundtrip latency number measure (especially your 1us)? Does zero copy imply mapping pages between processes? Is there an async kernel component involved (like I would infer from "io_uring") or just two user space processes mapping pages?
The actual (one-way cross-core) latency on modern CPUs varies by quite a lot [0], but a good rule of thumb is 100ns + 0.1ns per byte.
This measures the time for core A to write one or more cache lines to a shared memory region, and core B to read them. The latency is determined by the time it takes for the cache coherence protocol to transfer the cache lines between cores, which shows up as a number of L3 cache misses.
Interestingly, at the hardware level, in-process vs inter-process is irrelevant. What matters is the physical location of the cores which are communicating. This repo has some great visualizations and latency numbers for many different CPUs, as well as a benchmark you can run yourself:
Of course a round trip will take more time, but it’s not meaningfully different from two one-way transfers. You can just multiply the numbers I said by two. Generally it’s better to organize a system as a pipeline if you can though, rather than ping ponging cache lines back and forth doing a bunch of RPC.
I did read the original linux zerocopy papers from google for example, and at the time (when using tcp) the juice was worth the squeeze when payload was larger than than 10 kilobytes (or 20? Don’t remember right now and i’m on mobile).
Also a common technique is batching, so you amortise the round-trip time (this used to be the cost of sendmmsg/recvmmsg) over, say, 10 payloads.
So yeah that number alone can mean a lot or it can mean very little.
In my experience people that are doing low latency stuff already built their own thing around msg_zerocopy, io_uring and stuff :)
Little’s Law applies when the task monopolizes the time of the worker. When it is alternating between IO and compute, it can be off by a factor of two or more. And when it’s only considering IO, things get more muddled still.
some features are explicitly designed to minimize latency. I'm thinking of the IORING_SETUP_IOPOLL and IORING_SETUP_SQPOLL flags for io_uring_setup .
I'm not making that up, the manpage says that: https://manpages.debian.org/unstable/liburing-dev/io_uring_s...
Source: https://github.com/mvp-express/myra-transport/blob/main/benc...