Michael S (already5chosen.delete@this.yahoo.com) on May 30, 2022 3:12 am wrote:
> That is one of the cases where it should not be noticeable, at least when scalar JSON parser
> is written well. There is enough of other crap going on to make JSON parsing lost in noise.

I've measured ~6% of the CPU usage on a real server receiving JSON messages on parsing. Server is written in Javascript, so I'd imagine a more performant language may have a higher portion used by JSON parsing time.
JSON is chosen precisely due to its ease of use in the Javascript landscape (both client and server).

It's not exactly "large", but I wouldn't say "lost in noise" either. Faster JSON parsing would certainly be a benefit; wouldn't go out of my way for it, but if it's there, why not?

> On 100M$ farm wouldn't you lose more in most typical short JSON
> cases than you gained in hopefully rare long JSON case?

Really depends on your average load, and I'd imagine that SIMD helps in both long and short cases.

> And are there really so many 100M$ farms that consist completely of sort servers that do not clock
> themselves down when they see AVX512? I think, for majority of 1st and 2nd gen Xeon-SP SKUs it happend
> rather badly and this SKUs are still a majority of AVX512-capable server installed base.

Depends on whether you're thinking in the past or the future.
Also keep in mind the difference between AVX-512 FP and AVX-512 integer - the former throttles heavily, whilst the latter is much less (and IIRC doesn't throttle at all on Ice Lake Xeon).