Hacker News Reader

Comment by sanxiyn | Hacker News Reader

Acid2 pass itself maybe uninteresting, but Servo also does layout in parallel (right now, not in the future). No other engines do.

http://pcwalton.github.io/blog/2014/02/25/revamped-parallel-...

View on HN

Wow, Servo is way longer along than I thought!

From that post:

> No doubt about it, CSS 2.1 is tricky—floats perhaps more than anything else.

I really hope that Servo helps to identify parts of the HTML/CSS specs (if any) that unnecessarily prevent parallelism. By that I mean features (like "float" perhaps) that make it harder to parallelize, and where an alternative design could support the same use cases in a better way.

Those are the gems that teach us deep lessons about the problem space, and how future similar technologies out to be designed.

> I really hope that Servo helps to identify parts of the HTML/CSS specs (if any) that unnecessarily prevent parallelism.

This is beginning to happen; for example Servo work on running <iframe sandboxed> documents in parallel [1] led to the discovery that the HTML spec allowed different-origin iframes to mutate into same-origin, through the "document.domain" setter [2], which made process isolation infeasible (since it would require some way to merge separate processes with separate heaps back into a single process). That discussion led to a change in the spec [3].

[1] https://groups.google.com/forum/#!msg/mozilla.dev.servo/LQ46...

[2] http://lists.whatwg.org/htdig.cgi/whatwg-whatwg.org/2013-Aug...

[3] https://www.w3.org/Bugs/Public/show_bug.cgi?id=23040

CSS without float.. would not work very well! Unless we all go back to table based design (or ahead to flexbox I guess, I don't know enough, practically speaking, about that standard yet to reason proper about it though!)

If you clear your floats, you'll be fine in terms of giving the engine room to parallelize, according to pcwalton's blog post.

Does that 'allow' the classic use case of floats - potentially, what should be the only use of floats - wrapping text around an image? I feel like the answer would be "no", yet there's no good alternative for achieving this common requirement, which would cause a snag.

I believe you're right. But I don't think this is a huge problem: you simply don't get the benefits of layout out that specific text in parallel, it'll render sequentially as browsers tend to do today anyways. You can still get parallelization in other flows outside of the container of the floating elements.

I believe it does, yes. The point is that when you clear them later, the browser has a static guarantee that that float will not matter anymore. So the complexity due to the float is only relevant until the clear, and it can parallelize before and after the clear.

If you know you need to use floats to flow text around elements, feel free to use them! They're in CSS for a reason :) But if there are alternatives (such as absolute positioning) that work equally well in your particular design, then choosing an alternative may help you gain better performance.

Flexboxes don't really help. The reason that floats exist is so that elements can flow around them.

That's what they were designed for, but they are frequently used to do things that would better be done with flexboxes (if the support was there).

Wired.com made (web) history by being one of the first major sites to deploy a pure-CSS design, which was based on floats because that's all that worked back in the days of Netscape 4 and IE 4. The web design community now primarily uses floats for multi-column design because of its historical browser support. However if you don't need to support IE7 and lower, floats are one tool among many for multi-col layout, and not necessarily the best. Flexbox will only add another tool to that toolbox.

I imagine that is quite a huge feat! Here's a questions for the HN gurus though: how much of a browsers code deals with errors, compatibility, and legacy?? I imagine A LOT.

Fortunately, even the backward-compatibility quirks modes have a spec now: http://quirks.spec.whatwg.org/

Just having more and more of this stuff documented (instead of having to reverse-engineer it from existing browsers) makes things a lot easier than they once were.

Arguably a large part of the interest in Servo is whether the specs are good enough that developing a new browser from scratch is now feasible — certainly, yes, there is a lot of institutional knowledge around Mozilla, but a from-scratch implementation isn't going to exactly match Gecko everywhere.

Agreed 100%. We want to be part of the conversation as to how to make future CSS parallelizable and amenable to modern hardware, both to educate Web authors as to how to make fast content and to shape the work done in the relevant committees.

Does the lack of a multi-threaded layout engine really matter? Don't pages load fast enough already? And if layout performance really is becoming a problem, particularly in cases of dynamic DOM manipulation etc, isn't that largely the fault of the spec? Even shiny HTML5 is hardly designed for fiendishly complex UI layouts... it's an abomination for programmers and designers alike, and always has been. To this day it still requires herculean effort to achieve things that are trivial with real UI toolkits or publishing tools.

I guess now we're in this hole we best keep digging

HTML layout is not super-fast on low-end phones, and due to marketing stupidity those phones have 4-8 slow cores instead of 1-2 fast cores. I imagine Servo could do well there.

Drastically changing the HTML standard would take even longer than developing Servo.

> due to marketing stupidity those phones have 4-8 slow cores instead of 1-2 fast cores.

Beyond marketing, wouldn't more slow cores consume less battery power than fewer fast cores?

Yes, this is also part of Servo's strategy apparently: spread work across multiple cores, but each core does less and can be more efficient in terms of power draw. Faster + less battery use in one fell swoop. Though it's not really clear to me why this actually results in an efficiency win over maxing out a single core.

The way I've heard it, as a very imprecise rule of thumb, increasing clock speed scales power consumption quadratically, while adding CPUs increases power consumption linearly. So, very roughly speaking, one 1.6 GHz CPU uses twice as much power as two 800 MHz CPUs.

The equation for CMOS switching power (as opposed to leakage power) is

capacitance * voltage^2 * frequency

but the voltage limits the clock frequency. The exact scaling of maximum clock frequency with voltage depends on the circuit and the process, but from taking a glance at the voltage tables it looks like the main cores in current Snapdragons are generally running in the 0.8-1.2v range across their entire frequency range.

Most of the power differences between cores these days isn't in terms of clock speeds but in terms of extra structures that cause more instruction to be executed each cycle. The rule of thumb involving voltage/speed scaling is that your power use is indeed the square of your performance more or less when you're in a reasonable region. However, a relatively simple in-order core like an ARM A7 might also only take 1/4 the energy to execute a given instruction of a complex out-or-order core like an ARM A15 even when both are clocked at 1 GHz on the same process.

> Though it's not really clear to me why this actually results in an efficiency win over maxing out a single core.

i heard the whole multiple core thing is spurred in part by manufacturers trying to increase fab yield: if one of the cores on a multi-core chip is bad, it can still be sold, albeit at a lower price, rather than thrown out.

If apps were parallel, yes. But most of them aren't.

If Servo works, most* of the Web will be parallel.

* This is probably too optimistic.

I'd like to see a memory usage comparison with Servo, since mobile devices have far less memory than desktops.

On the other hand, I believe that HTML layout can be done, using only a single thread, much faster and with far less memory than current mainstream browsers do, by greatly simplifying the code (i.e. remove excessive use of abstraction, using different data structures, etc.) Not exactly the same, but along the same lines as, this related item that appeared here a few days ago: https://news.ycombinator.com/item?id=7457674

You'd be surprised: it's not that easy to beat existing browser engines. The IE trick you mention is pretty trivial. That said, it is possible to win, and one of the great benefits of Servo is that its data structures are simpler than other browser engines' in many ways. But it's not as easy as you think.

I think that it's clear that any technology that helps the browser better compete with native applications in terms of performance is worthwhile.

Regarding new specs, we're of course working on that too, with specifications like flex box and grid layout. Part of Servo's goal is to steer the conversation toward what can be done to make future CSS specs parallelizable on CPUs and GPUs. But of course we want to be fast on existing Web content as well.

Native applications do layout and engines such as Qt are single threaded.

Why is that a reason to remain single-threaded? Especially when most new hardware is multi-core.

Consider the parallel addition task, it takes N*log(N) instead of single threaded addition which takes N. Which is more energy efficient?

If takes 10x the power to do layout in 20ms vs 30ms is it still a good deal?

10x is a silly exaggeration. As for your general point, you'd be surprised: often consuming more power per second is much better for power efficiency than consuming less power if that power consumption lets you complete the task faster. That is because the wins of getting back to the CPU's low-power idle state quicker dominate everything else. It is almost always better to use the power available to you.

Try drawing lager pages. That is the bottleneck in one of my applications.

> Don't pages load fast enough already?

Ho ho ho... good one.

why all the effort to build the flow tree? can't any block be rendered in parallel to a buffer and then copied into the correct location once all the widths and heights are known? in the examples given it seems like all blocks including the green one could be rendered in parallel to separate buffers.

Layout is precisely about learning all the widths and heights. No rendering whatsoever happens in layout. And as you pointed out, rendering can be parallelized rather easily.

Reply on news.ycombinator.com