←back to thread

io_uring is faster than mmap

(www.bitflux.ai)
283 points ghuntley | 3 comments | 04 Sep 25 22:01 UTC | HN request time: 0s | source
Show context
hsn915 ◴[04 Sep 25 23:56 UTC] No.45133595[source]▶
Shouldn't this be "io_uring is faster than mmap"?

I guess that would not get much engagement though!

That said, cool write up and experiment.

replies(3): >>45133612 #>>45134589 #>>45136289 #
nine_k ◴[05 Sep 25 08:29 UTC] No.45136289[source]▶
No. "io_uring faster than mmap" is sort of a truism: sequential page faults are slower than carefully orchestrated async I/O. The point of the article is that reading directly from a PCIe device, such as an NVMe flash, can actually be faster than caching things in RAM first.
replies(1): >>45140950 #
wmf ◴[05 Sep 25 17:11 UTC] No.45140950[source]▶
reading directly from a PCIe device, such as an NVMe flash, can actually be faster than caching things in RAM first.

That's not true though, because the PCIe device DMAs into RAM anyway.

replies(1): >>45145587 #
1. nine_k ◴[06 Sep 25 01:02 UTC] No.45145587[source]▶
No, it can DMA straight into L3 cache, as mentioned in the article.

See https://www.intel.com/content/www/us/en/io/data-direct-i-o-t...

replies(1): >>45146256 #
2. wmf ◴[06 Sep 25 03:01 UTC] No.45146256[source]▶
Your server doesn't have DDIO turned on though.
replies(1): >>45146577 #
3. nine_k ◴[06 Sep 25 04:19 UTC] No.45146577[source]▶
Not mine, but fair!