32 comments

[ 3.9 ms ] story [ 89.2 ms ] thread
Wait, that's like 4.4*10^7 clock cycles for my laptop!

Here's hello.sh (in bash) on a Raspberry Pi 3B+:

  #!/bin/bash
  echo "Hello World!"



  xxxxx@xxxxxxx:~ $ time ./hello.sh
  Hello World!

  real    0m0.010s
  user    0m0.001s
  sys     0m0.010s

Yay, I beat the high score!
Holy shit, that’s sad
It's a startup time of the runtime. More interested in results with two "hello,world"s. How much the next one going to cost.

Also C++ with cout, -O3 is 4ms and one in *.Zig is 1ms for some reason.

Probably want -Os (and strip afterwards) if you're interested in startup time. C is also 1ms. C++ has a big cxxrt that zig doesn't. FWIW c++ only takes 2ms on my (quite slow) machine.
would ios_base::sync_with_stdio(false); also help?
In my benchmark, the c++ version uses the c api (printf rather than cout), so it wouldn't make a difference.
Yup. Dart has a runtime. Runtimes have to spin up when they're run cold.

But we shouldn't let facts get in the way of a good Hackernews hivemind dogpile. Dart bad!

Both process startup and terminal can be very slow depending on OS.

I also have a rule, I don't bash novice developers for their lack of knowledge.

(comment deleted)
Do you really think novice devs work on Dart at Google?
Would that be really so surprising?
How it compares with go, node and java?
Here are a few comparisons (13" MBP, YMMV).

  $ ./hyperfine './hello.gonative' --warmup 10
  Benchmark #1: ./hello.gonative
    Time (mean ± σ):       3.3 ms ±   0.9 ms   [User: 1.2 ms, 
      System: 1.2 ms]
    Range (min … max):     2.4 ms …  14.2 ms   455 runs

  $ ./hyperfine './hello.dartnative' --warmup 10
  Benchmark #1: ./hello.dartnative
    Time (mean ± σ):       8.2 ms ±   0.4 ms   [User: 3.9 ms, 
      System: 3.6 ms]
    Range (min … max):     7.5 ms …  10.2 ms   243 runs

  $ hyperfine 'java HelloWorld' --warmup 10
  Benchmark #1: java HelloWorld
    Time (mean ± σ):      90.7 ms ±   1.3 ms   [User: 83.2 ms, 
      System: 23.9 ms]
    Range (min … max):    88.7 ms …  94.2 ms   31 runs
Which JDK? OpenJDK 14 on my machine:

    ~> Measure-Command { java Main }
    Milliseconds      : 59
Also with GraalVM Native:

   ~> .\native-image.cmd --static Main Main
   ~> Measure-Command { ./Main }
   Milliseconds      : 16
I got:

OpenJDK 8: ~80-90 ms

OpenJDK 11: ~60 ms

OpenJDK 14: ~50-55 ms

The first runs were a bit slower, perhaps because the files needed by the JVM weren't in disk cache or something.

Other languages/runtimes:

Python 2.7: 20-25 ms

Python 3.8: 30-40 ms

Native binary compiled from C: 1-3 ms (so bordering on unmeasurable at millisecond resolution)

This was on Linux, and an entirely unrigorous experiment.

edit: formatting

Nice! I'm glad startup time is getting closer to python. FWIW GraalVM native image was 16ms for me on Windows.
Lots of bashing in this thread, but to clarify a bit:

The Dart VM is roughly based on V8's internals. Running dart alone would be similar to running node, which needs to initialize the VM, parse the code, and run it. This is obviously slower than a pre-compiled binary.

dart2native brings it half-way to native, by doing parsing and initialization, and saving a memory snapshot right before executing the main function. This greatly reduce startup time (and is the same technique that makes Flutter apps start fast), but still not as fast as a binary, since it is loading the snapshots and doing some initialization.

This is usually fast enough, because once the initialization is ran, the following code will run fast. This is a constant initialization loading time, and doesn't represent anything.

Saying 22ms doesn't mean anything, and feels like a misinformed statement.

Dart overall is a pleasant language, and it is somewhat unique that it provides multiple ways to run. You can run with dart, dart2native, dart2js, and I believe there's one for ARM (for Flutter). This allows you to choose the best for your target platform

As people pointed out in the thread, in 22ms a game will have rendered a full 3D world with physics and audio multiple times.

22ms is... not fast. Unless you count modern world where "fast" is "anything that runs under 10 seconds".

That's an odd contrast (one frame of a game versus loading a language runtime). How long does a game engine take to load?

A reasonable comparison is whether Dart is able to render complex experiences at 60fps (i.e. 16ms / frame) on a low-end phone device, which is the primary use case.

Depends on the game engine. Jonathan Blow has written one for his upcoming Sokoban game and it completely compiles and launches the game from source in about a second and a half. No incremental compilation, no debug mode. That's not just the runtime engine, it's the tooling to make the game, as well. A second and a half, from source files on disk to a running binary showing a 3D viewport on screen.

Think about that.

There is no excuse for how slow software is today. None.

Many parts of Windows 95 were faster in wall clock time in 1995 on the hardware of 1995 than today's Windows 10 is on the hardware of today. Yes, today's software does more, but THAT MUCH more? Are you sure?

The hardware we have is very fast. Software developers have been relying on hardware upgrades for performance improvements for far too long, and now few software developers know how fast things can be, if they just try just a tiny little bit.

Also, OOP teaches developers how to think about software in ways that are exactly opposite to how computers actually do work efficiently. Object oriented programming is just inherently slower because it encourages developers to think of things one at a time. Computers like to do things in batches.

More people need to think about performance, because clock speeds aren't going up like they used to, and we still don't know how to write software that spreads across a lot of cores very well. The free ride that hardware upgrades provided us is quickly coming to an end.

tl:dr; everyone needs to learn about how processor caches work, especially the 24-year old JS devs who think they already know everything.

Studies have shown that user interface latency was at its all time best in the mid-80’s, assuming the user was using a dumb terminal that talked to a mainframe.

Modern software sucks.

This is like saying a game runs at 0.1fps because it took a couple seconds to load at the start.

This is initial load time. Printing likely account for <1% of the actual time, the rest being initializing the program

Other examples in the thread show people routinely outputting "Hello, world" from cold start in under 5ms.
All that proves is that V8 is a beast, similar to the JVM. Here's what an alternative world could look like:

  $ time luajit -e 'print"hello world"'
  hello world

  real 0m0.002s
  user 0m0.000s
  sys  0m0.002s
But it's worse than that, because the time at issue is for the AOT-compiled native code binary.

To be fair, I presume that the included runtime is written in C++, and modern C++ coding styles are pretty inefficient. Lots of small, dynamic allocations. Too much of the wrong kinds of abstraction. It's what happens when you shoehorn problems into established solution models, rather than approach development, including boiler plate code, as crafting a bespoke solution model fit to the problem. But the latter makes onboarding new engineers more difficult--you can't start hacking without first grasping the larger architecture--and therefore frowned upon in larger companies, or in small companies that equivocate big company approaches to "best practice".

Poking around on my laptop, dash and perl take about 2ms. Python takes 12.

Anyway, the modern c++ I use avoids allocations like the plague.

I subscribe to the idea that C++ is a language for standard library implementors, and that std:: is mostly placeholders for the “real” implementation you’ll use later.

[disclaimer: I work on the Dart team.]

As others have already noted, the intent here is to purely measure the one-time startup cost of the runtime in ahead-of-time compiled code - i.e., the amount of time necessary before we can starting executing any user code - i.e., the print in this case. It's assumed that the time to do the actual print is trivial here.

If you want to compare, you might compare with other garbage collected language/runtimes.

The text on the page is "AOT-compile apps to native machine code for instant startup".

And now you're saying that the start up is actually not instant.

- LuaJIT: 0.002ms [1]

- Dash and perl take about 2ms. Python takes 12. [2]

- Perl 0.008s [3]

- Python 0.009s [4]

And that's before we talk about compilation speeds [5]

[1] https://news.ycombinator.com/item?id=23109404

[2] https://news.ycombinator.com/item?id=23110092

[3] https://twitter.com/beeonbird/status/1258472898446180353

[4] https://twitter.com/f00zz_/status/1258427236891480067

[5] https://twitter.com/alexyanov/status/1258474918607544322

"Instant" is a relative word (unless we're actually at 0 time :-)).

We're focus on client apps, where we're really focused on human perception. In that sense, I'd consider 22ms or 2ms as instant (for startup). I wouldn't consider 2s as instant.

I'm not saying Dart doesn't have room to get better, but there are diminishing returns at a certain point.

22ms here. 22ms there. "Focused on human perception" here, "Focused on human perception" there.

And suddenly we have what is basically a supercomputer unable to perform anything without lag.