[arturadib]

Oh look another JS vs native benchmark...

I still think we're missing the point; we're focusing on the wrong benchmarks. We're putting too much emphasis on JS speed, when in reality most web apps are behind native because of lack of (perceived) graphics performance.

The shortcut out of this requires a greater variety of GPU-accelerated CSS transitions/DOM changes, as well as easier computations and DOM manipulations off the main thread, which cause horribly noticeable UI hiccups. Web Workers are still too primitive (e.g. no DOM access whatsoever) and slow (e.g. no shared memory).

Not saying it's unimportant to improve JS's CPU performance; just saying that we're focusing too much on the wrong battle in the war against native.

[SigmundA]

WebGL anyone? This is what does most of the heavy lifting in the main asm.js use case: games.

I am still waiting to see a UI framework built using asm.js and WebGL. Maybe even a existing native framework ported over, something like WPFe(Silverlight) might be about as difficult as a game engine to port.

Asm.js still has no solution to shared memory threads, which will be an issue for any modern game engine I would imagine.

[mistercow]

>Asm.js still has no solution to shared memory threads, which will be an issue for any modern game engine I would imagine.

This feels to me like the elephant in the room, not just for asm.js, but for JS in general. We have Web Workers, and that sounds like a solution, so it dampened the discussion of parallelism on the web. But in a lot of situations, Web Workers just don't cut it. Copying memory back and forth is just too inefficient.

[SigmundA]

I agree, I can appreciate the idea of share nothing threading(web workers) which is more like multi-process rather than multi-threading, it's probably the better model for much of the threading use cases out there. However there are performance critical scenarios where lightweight shared memory threads will always win and will be required if it is to be a valid compile target for c/c++ coded bases.

[nitrogen]

If there's a way to do zero-copy message passing in web workers (or equivalently, overhead-free transfer of object ownership between threads), shared memory may be unnecessary.

[gmac]

This is available in recent versions of Chrome and Firefox, on the form of Transferable objects: https://developer.mozilla.org/en-US/docs/Web/API/Transferabl...

[mistercow]

That helps, but it still means you can only have the data on one thread at a time. What would be a massive help would be if you could split a buffer into non-overlapping views, and transfer each view to a separate worker. Some algorithms would still be challenging or impossible to implement this way, like parallel prefix sum algorithms, but it would still greatly widen the number of things you could do in parallel.

[ris]

Doing such a thing would implicitly require doing shared memory behind the scenes.

[nitrogen]

Yes, at some layer of abstraction there will be the possibility of sharing memory. It's probably better to keep it behind the scenes for all but the most CPU intensive uses.

[sitkack]

The js app performance will be a "solved" problem in 4 years.

* WebCL will be widely deployed, better penetration than current WebGL

* Web workers will be able to share typed arrays by reference (see Mozilla's Transferable Objects)

* Other vectorization libraries will be available (Rivertrail)

   http://blog.aventine.se/post/16318162396/simd
   http://www.khronos.org/webcl/
   https://github.com/RiverTrail/RiverTrail

[jnbiche]

Why do you think WebCL will be widely-deployed in 4 years? I used to think that, not I'm not so sure. It's sure not been given a warm welcome by any of the major browsers. I'm not sure I share your optimism.

I do agree that Transferable Objects will be a tremendous boon, and should address all the shared memory thread complaints elsewhere in these comments.

I had forgotten about RiverTrail. Very cool project, but is it even still in active development? Commits seemed to have trailed off.

[sitkack]

heh, trailed off.

----

Because WebGL will be based off of WebCL, WebGL is a subset targeted at graphics.

And change is accelerating, everything in the future happens faster. So there will be more people to implement WebCL than there was to implement WebGL. More and more people are using advanced web apps so the pressure is higher to make things faster.

[andhow]

The 'war on native' has multiple fronts and this post is just reporting on one of them. For many apps, I agree that the items you've mentioned are the most significant and progress is also being made on these fronts (e.g., GPU-acceleration of some CSS transitions are already in Firefox; more and more DOM APIs are being exposed to workers). For other apps, though, especially ones using WebGL like games or authoring tools, raw computational throughput is the most significant.

[belluchan]

Web workers will never have DOM access. They run in a separate process with a separate URL. That's not an issue. You use postMessage to communicate with the workers.

[goldenkey]

I'm pretty sure OP knows about postMessage. That wasn't his point.

The only way to pass data is through message passing - but the message passing takes only strings. Just object serialization and deserialization in JS is slow enough to make it infeasable to do large-scale transfer between the two without paying for transmuting the data.

We need immutable object passing for performance. Or locked mutable object passing (like Rust.)

And why not the ability to modify DOM elements with the repaints happening on the main thread?

[woogley]

You're looking for Transferable Objects:

https://developer.mozilla.org/en-US/docs/Web/Guide/Performan...

[chongli]

Still seems like quite a hack. Why can't we just have immutable objects shared by all webworkers?

[goldenkey]

Sweet beavus!

[marcosscriven]

I had exactly this issue when trying to make my OpenSCAD port multithreaded in asm.js - the results of each worker took longer to serialise/deserialise than the time saved by doing them in parallel. Unfortunately couldn't even use transferable objects, as the C++ objects in question were CGAL nef polys, and once copied across would be garbage. It would work though if you have objects you can allocate to a specific buffer (I since discovered called a 'bump allocator')

https://groups.google.com/forum/m/#!searchin/emscripten-disc...

[mistercow]

>And why not the ability to modify DOM elements with the repaints happening on the main thread?

It's not just a matter of repaints. You'd actually have to make the DOM thread safe, which is not trivial.

[goldenkey]

I think locking the DOM might not be too difficult. To have multi-threaded accessors/getters would be insanely tough given the size of the existing code-base. But a single exclusive lock over the DOM could work.

[mistercow]

There are several problem with that. One is that the web worker would block all user input, including scrolling, while you were modifying the DOM. Also, the locking would absolutely have to be manual. Locking before each DOM modification and then unlocking after would be disastrous, since you'd have no guarantees about the DOM's state from one call to the next. And once you introduce manual locking, you're largely defeating all of the safety guaranteed by workers in the first place.

Or think of it this way: If you can lock the DOM reliably and efficiently, then you can write code that shares memory by storing data in the DOM, rather than copying it between workers. That would get you most of what you get from ordinary threads, albeit with only one lock available. This would be such a huge convenience and performance win, that if it were possible, they'd have done it to allow memory sharing without going through the DOM at all. That right there should tell you that allowing DOM access from workers by locking is impractical.