## Wednesday, August 17, 2011

### The Lesson V8 Can Teach Python and Other Dynamic Languages

Being unable to completely give up math for computers, I am naturally drawn to Project Euler and as a result solved a ridiculous number of the problems posted there while learning Python. A few months ago (March 13), after reading Secrets of a Javascript Ninja, I decided to begin converting my solutions to Javascript. A month and a half later I came back to it, and then finally two months after that, I began to take it seriously.

After making this decision, I noticed the prime Sieve of Eratosthenes was mighty fast when I ran it in Chrome, maybe even faster than my beloved Python. I tabled the thought for a little, but never really forgot it. So a few weeks ago (early August 2011), I finally got a working install of node running on my machine and was able to make more of this thought. (I say finally installed because on two previous tries I gave up because of conflicts with my version of gcc, coupled with the fact that I had no good reason to use node.)

When I originally did the conversion, I had skipped problem 8, because my implementation required pulling in the problem data as text from a file. While hanging out with Boris and Eric from on the Chrome Developer Relations team, I decided to give it another go on node (xhr requests not allowed) and found it to be quite simple with readFileSync in the node native fs module. After witnessing this, over this weekend, I decided to harness the power of V8 -- the Javascript engine that powers Chrome and node -- and run all my scripts locally with node. So over a two day period, I hack-hack-hacked my way into converting the Python solutions for problems 11 through 50 (the remaining unconverted) into their Javascript equivalents, while also converting a good portion of my hefty functions module.

Once this was done, I had also found I could replace most of the nice parts about Python with my own equivalent. For example, I was able to replace functionality I needed from the Python set datatype with
function uniq(arr) {
var result = {};
for (var i = 0, val; val = arr[i]; i++) {
result[val] = true;
}

return Object.keys(result);
};

and I was able to replace the (amazingly) useful Python handling of long integers with a non-native node package called bigint that uses libgmp among other usings. Of course, for Python's secret sauce -- the list comprehension -- I was able to substitute enough filter, reduce and map statements to almost make it seem like I had never left Pythonland. After doing all this, I also ended up writing my own operator.js to replace the wonderful Python native module operator, and my own timer.js to stand in for the Python native module time.

Finally, I had working code and could do a side by side comparison of V8 and the Python interpreter. Update: I added a column for PyPy, a just in time implementation of Python. Here is what I found (averaging the runtime over 10 separate calls to each function, the results are):

1*2331681804ms1215ms1.48385ms
2*4613732247ms102ms2.4285ms
3*68574725ms1508ms3.13582ms
49066098708ms149ms58.44282ms
5*232792560136ms186ms0.73114ms
6*2516415010ms4ms2.506ms
7104743656ms12ms54.6711ms
8*4082418045ms5014ms3.607042ms
931875000610ms3ms203.338ms
101429138289226628ms167ms39.69116ms
117060067449ms2ms24.5011ms
12765765005127ms203ms25.26100ms
13*55373762301795ms10710ms0.171423ms
148377995572ms1712ms3.25362ms
15*13784652882054ms18ms3.0055ms
16*13661844ms265ms6.96462ms
172112487ms4ms21.757ms
18*10742291ms1790ms1.281090ms
19*1712254ms336ms6.71342ms
20*6481061ms9154ms0.12374ms
213162618910ms1038ms18.22728ms
22871198282188ms7ms26.868ms
23417987183318ms1120ms74.391295ms
24*2783915460206ms210ms0.98139ms
2547825865ms35ms167.57232ms
2698328ms18ms1.564ms
27-59231645738ms22536ms28.6528288ms
28*6691710018509ms1037ms8.21981ms
299183184ms96ms1.9220ms
3044383952167ms1037ms50.31877ms
31736829606ms257ms37.38154ms
3245228206888ms12096ms17.104266ms
33100300ms6ms50.0015ms
34407307462ms2447ms3.05247ms
35558617ms848ms10.16242ms
36872187189788ms2183ms86.943532ms
377483172389022ms71845ms33.2561551ms
38932718654506ms10ms50.6012ms
39840178ms6ms29.6712ms
40*210326ms202ms1.61119ms
4176524132627ms133ms19.7565ms
4216265ms7ms9.298ms
431669533489038ms2ms19.002ms
445482660384013ms27744ms13.846621ms
45*153377680517ms4ms4.258ms
4657772864ms202ms14.1865ms
47134043400967ms12838ms31.234425ms
48911084670046ms16ms2.886ms
49296962999629115ms8ms14.3813ms
509976513277ms80ms40.9651ms
*These were very quick to run, so the runtimes are the time taken to run 10000 times.

As you'll notice, standard Python gets its butt kicked. I was kind of saddened by this, but in the end, just giddy that our web is faster because of it (90% of my life is digital) and also that we can do scripting faster on the server side (attribute to Boris Smus) because of the node project (thanks Ryan Dahl).

Standard Python is actually slower in 46 of the 50 problems. In 28 of the 46 node is faster by a factor of 10 or greater, in 9 of those 28 by a factor of 50 or greater and in 2 of the 9 by a factor of 100 or greater! The only 4 in which Python was faster were from the n = 10000 sample. In fact, I was able to pinpoint exactly why:
• #5 - My own Javascript implementation of gcd is slower than the native (from fractions import gcd) Python library (resulting in a difference of 50 ms over 10000 iterations)
• #13 - The node package bigint is slower than the Python native long int (Javascript is slower by a factor of 6)
• #20 - The node package bigint is slower than the Python native long int (Javascript is slower by a factor of 8.5)
• #24 - Having to perform two slices is slower in Javascript than in Python and there is no good way to just remove one element (resulting in a difference of 4 ms over 10000 iterations; a little bit about that here)
So what, you ask, is that lesson I speak of? Well, Javascript didn't used to be this fast. How did it get that way? The brilliant and inspired people behind V8 rethought the Javascript compile steps and after much work, we now have code that is closer to the metal (attribute to: Eric Bidelman, i.e. closer to machine code) than we had ever had before. The use of just-in-time compilation and other incredible techniques has taken a formerly slow and clunky language (Javascript) which was used well beyond its original intended scope, and turned it into a damn fast dynamic language. Hopefully, this movement will make its way to Python and other dynamic languages and we can all have our code end up this close to the metal.

Update: In response to the comments, I ran the same code on the same machine, but with PyPy in place of Python. This is the direction I hope standard Python goes in and commend the guys pumping out faster and faster just in time implementations over at PyPy. I went through and counted 20 node wins, 29 PyPy wins and 1 tie. (I reserve the right to update the post with more detailed metrics.) While I do commend them, the results don't really belong in this post because PyPy is still an offshoot. (However, as I understand, they both have ties to C, as PyPy uses GCC to compile to bytecode and V8 is written in C++. Feel free to supplement my knowledge in the comments.)

Update: All benchmarking was run on my Mac Pro Desktop with a 3.2 GHz Quad-Core Intel Xeon processor and 4 cores for a total of 12 GB 1066 MHz DDR3 memory. I used Python version 2.6.1, node version 0.4.9, and PyPy version 1.5 (running on top of Python 2.7.1 with GCC 4.0.1).