Fork me on GitHub
We are hiring a PhD student to do research on functional high performance computing, including work on Futhark itself. See here for more details.

Futhark with Fangs!

Posted on April 1, 2018 by Troels Henriksen

Any fledgling language designer must consider how to make their language useful to someone, and thus gain users. Last year, the Futhark outreach efforts were focused on catering to an overlooked niche - computational history - but this niche may have been just a tad too narrow. In fact, I am still not sure whether it exists at all. Therefore, this year, we will change our strategy to focus on the most mainstream development field in the world: web programming. This field is characterised by two key properties:

This analysis inspired me to implement support for calling Futhark functions via HTTP. Futhark already has a pretty good Python interface, so it was a quick hack to wire it up to the standard http.server module. The result: Futhark with Fangs.

Using Futhark with Fangs

Putting Futhark on the web is quite easy. First we define a couple of interesting Futhark functions in a file futapp.fut:

entry dotprod (xs: []i32) (ys: []i32) = reduce (+) 0 (map2 (+) xs ys)

entry sumrows (xss: [][]i32) = map (reduce (+) 0) xss

We then compile it to a Futhark library with embedded GPU code:

$ futhark-pyopencl --library futapp.fut

And then we simply pass the module name to Futhark with Fangs:

$ ./futhark_with_fangs.py futapp

We can now invoke the Futhark entry points using our favourite HTTP client:

$ echo '[1,2] [3,4]' | curl -X POST --data-binary @- localhost:8000/dotprod
10i32
$ echo '[[1,2], [3,4]]' | curl -X POST --data-binary @- localhost:8000/sumrows
[3i32, 7i32]

The input and output values are represented in the soon-to-be industry standard Futhark value notation. However, we can also pass the input in a somewhat more efficient binary data format. It is a little tricky to write in a shell, so we will generate some random data with futhark-dataset:

$ futhark-dataset --binary -g [2][2]i32 | curl -X POST --data-binary @- localhost:8001/sumrows
[-807407905i32, 1953530750i32]

Transmitting large arrays over HTTP is of course rather slow. A good idea would therefore be to use compression to cut the transmission times. So, does Futhark with Fangs support HTTP compression with something like gzip? I don’t know, as I did not read the http.server docs closely enough. Maybe it does; we will just have to wait and see. This seems in the spirit of dynamically typed programming languages.

Functional programmers, of which I am one, care a lot about laws. One of these, Wirth's Law, states roughly:

“Software is getting slower more rapidly than hardware is getting faster.”

– Niklaus Wirth

GPUs, with their tremendous computational performance, are already terribly close to breaking this law, and I would rather not see Futhark become accessory to a crime. Compiling Futhark to Python is a valiant effort to reign in the performance a bit, but it still runs quite fast, as most of the work remains on the GPU. Hobbling Futhark with HTTP may be just what we need to obtain compliance with Wirth’s Law.

Is this a joke?

Check the date of the post. However, do note that Futhark with Fangs is fully operational and works quite well for what it is. Perhaps Stockholm Syndrome is setting in, but it even feels like it might actually be useful in some cases. For problems that contain a large ratio of computation to I/O, which includes most Monte Carlo simulations, the network overhead may be negligible. So at this point, I’m not even sure myself whether Futhark on Fangs is supposed to be a prank. I guess the joke is on me.