News for hackers. Stuff that matters.
Re:Fast function to parse strings into double (binary64) flo... by scottlamb on Tuesday March 10, 2020 @07:52PM (#22541356)
It has about 15 KB of data: 10KB in power_of_ten_components and 5KB in mantissa_128. [1] That's ~23% of the per-core L1 data cache on (eg) the recently announced APU chip. [2, 3] So before getting too excited about the microbenchmark numbers, I'd be careful to ensure the additional cache pressure doesn't slow down my program as a whole.

edit: also, ~2.5KB of that is due to padding in power_of_ten_components. I wonder if it'd be better to split the uint64_t field into two uint32_ts to avoid this.

[1] https://github.com/lemire/fast_double_parser/blob/master/inc... [2] https://news.ycombinator.com/item?id=22440894 [3] https://en.wikichip.org/wiki/amd/ryzen_embedded/r1606g

Re:Fast function to parse strings into double (binary64) flo... by userbinator on Tuesday March 10, 2020 @08:19PM (#22541551)
Also, I know it's not always possible, but if your application is spending so much time parsing strings into numbers that this could make a difference, seriously consider whether the source of those strings could be changed to not emit strings, because unless it's something like user input, chances are they may already be doubles which can be used as-is.
Re:Fast function to parse strings into double (binary64) flo... by thedance on Tuesday March 10, 2020 @10:29PM (#22542376)
HN is collectively schizophrenic on this topic. On the one hand, the comments say that nobody really needs anything fancier than JSON. On the other hand, other comments say that if JSON parsing shows up in your profiles, you should switch off JSON!
Re:Fast function to parse strings into double (binary64) flo... by scottlamb on Tuesday March 10, 2020 @10:49PM (#22542492)
I think if you're a little more generous and say "you rarely need" rather than "nobody really needs", there's no contradiction. Most of the time, you're not using much CPU in total or your server is doing something much heavier than parsing and serialization. Either way, JSON is fine. Improving its performance is nice but not worth a ton of effort. Once in a while, it's significant, and you switch to a binary format if you can.

And like anything, it depends on what you're doing. If you're doing lots of server-to-server RPCs on big distributed systems like within Google, it's much more reasonable to start with protobuf or the like than with JSON. And not just for performance reasons; having the generated code helps a lot with maintainability. (Some JSON libraries have something similar; some don't.)

Re:Fast function to parse strings into double (binary64) flo... by zozbot234 on Wednesday March 11, 2020 @10:51AM (#22546289)
Hexfloats are not a binary format, they're as text based as anything else. And they get rid of significant bottlenecks in float parsing.
Re:Fast function to parse strings into double (binary64) flo... by thedance on Wednesday March 11, 2020 @12:55PM (#22547367)
But the existence of the hex float format, and the fact that it accepts either 0x or 0X, upper or lowercase characters for A-F, dot or comma for the point depending on locale, and either p or P for the exponent, all contribute to the reasons why strtod is so slow. The faster library in this article handles none of those things. It also does not handle NaN or INF.
Re:Fast function to parse strings into double (binary64) flo... by userbinator on Tuesday March 10, 2020 @10:47PM (#22542479)
HN (and a lot of other places, really) has a large population of web or otherwise "web-heritage" developers, as well as those whose ventures into computing started beside/before the web. I suspect the JSON promotion comes mostly from the former group, while the latter, of which I am a member, laments it roughly as much as we did the "XML craze" of the late 90s and early 2000s.
Re:Fast function to parse strings into double (binary64) flo... by xxs on Wednesday March 11, 2020 @08:59AM (#22545409)
Indeed, hand rolled fixed point network marshaling, preferably with delta encoding for time series alike data (or data that oscillates) is couple of orders of magnitude faster and but also significantly "smaller" than a double-string-double, counterpart. Getting ~1 byte per data point of time series is quite achievable.
Re:Fast function to parse strings into double (binary64) flo... by vbezhenar on Wednesday March 11, 2020 @09:17AM (#22545545)
How do you encode floating point delta into one byte?
Re:Fast function to parse strings into double (binary64) flo... by thedance on Wednesday March 11, 2020 @12:56PM (#22547375)
You can't, that's why they said fixed point. Time series data should probably never use floats, for the exact reason that delta-encoding is not effective.
Re:Fast function to parse strings into double (binary64) flo... by scottlamb on Wednesday March 11, 2020 @05:50PM (#22551173)
"Can't" is too strong a word. Depending on what your numbers represent, I'd think you could do one or more of:

* have a shorter exponent and/or mantissa to make each number shorter and the overall series more repetitive. (maybe even skip the mantissa if you only care about order of magnitude! call it the opposite of fixed-point, or just storing the logarithm of the value, whatever terminology works for you...)

* skip the sign bit for a non-negative series (e.g., deltas of a non-decreasing series),

* use a variable-width bitstream encoding like exp-Golomb to represent the exponent and/or mantissa,

* use run-length encoding if it's repetitive (after delta-encoding),

* etc.

One byte per value (or less!) is surely possible in some cases. But I've never done it myself and agree fixed-point sounds much more pleasant. One of my projects represents deltas of durations in units of 90,000ths of a second, varint-encoded. I much prefer that to dealing with float seconds.

Re:Fast function to parse strings into double (binary64) flo... by thedance on Wednesday March 11, 2020 @07:40PM (#22551988)
There's also the XOR compression scheme described by Facebook:

https://www.vldb.org/pvldb/vol8/p1816-teller.pdf

Re:Fast function to parse strings into double (binary64) flo... by cm2187 on Wednesday March 11, 2020 @05:05AM (#22544250)
And a tiny minority I am proudly part of, who watches these debates about xml vs json or tabs vs whitespace with the same amusement than Mike Judge!
Re:Fast function to parse strings into double (binary64) flo... by xxs on Wednesday March 11, 2020 @09:00AM (#22545417)
C'mon spaces vs tab is a serious one and it makes sense... unlike xml, json shizz.
Re:Fast function to parse strings into double (binary64) flo... by saagarjha on Wednesday March 11, 2020 @09:02AM (#22545434)
Spaces versus tabs is a stupid debate because "tabs for indentation, spaces for alignment" is clearly the correct answer.
Re:Fast function to parse strings into double (binary64) flo... by XMPPwocky on Tuesday March 10, 2020 @10:55PM (#22542523)
Couldn't it just be different users posting those?
Re:Fast function to parse strings into double (binary64) flo... by mandelbrotwurst on Wednesday March 11, 2020 @12:35AM (#22543022)
That's exactly what this person is saying / what they intended by the phrase "collectively schizophrenic".
Re:Fast function to parse strings into double (binary64) flo... by XMPPwocky on Wednesday March 11, 2020 @12:55AM (#22543113)
Okay, but then "people disagree about this" seems a bit less interesting?
Re:Fast function to parse strings into double (binary64) flo... by gameswithgo on Tuesday March 10, 2020 @10:40PM (#22542444)
that just reflects that actual engineers and not actual engineers frequent hn
Re:Fast function to parse strings into double (binary64) flo... by kzrdude on Wednesday March 11, 2020 @05:04AM (#22544244)
A serious question is how can we make microbenchmarks that take this into account? Along the lines of measuring a speedup of an operation, in varying scenarios of cache pressure.
Re:Fast function to parse strings into double (binary64) flo... by vbezhenar on Wednesday March 11, 2020 @09:21AM (#22545570)
I'm not sure about actual implementation, but one approach would be to copy this function 100 times and invoke those copies sequentially. Of course linker must not optimize it into one function, so that's why I'm not sure about actual implementation. May be naive approach would work or may be not.