I am trying to choose the right technology to use for updating a project that basically renders thousands of points in a zoomable, pannable graph. The current implementation, using Protovis, is underperformant. Check it out here:
http://www.planethunters.org/classify
There are about 2000 points when fully zoomed out. Try using the handles on the bottom to zoom in a bit, and drag it to pan around. You will see that it is quite choppy and your CPU usage probably goes up to 100% on one core unless you have a really fast computer. Each change to the focus area calls a redraw to protovis which is pretty darn slow and is worse with more points drawn.
I would like to make some updates to the interface as well as change the underlying visualization technology to be more responsive with animation and interaction. From the following article, it seems like the choice is between another SVG-based library, or a canvas-based one:
http://www.sitepoint.com/how-to-choose-between-canvas-and-svg/
d3.js, which grew out of Protovis, is SVG-based and is supposed to be better at rendering animations. However, I’m dubious as to how much better and what its performance ceiling is. For that reason, I’m also considering a more complete overhaul using a canvas-based library like KineticJS. However, before I get too far into using one approach or another, I’d like to hear from someone who has done a similar web application with this much data and get their opinion.
The most important thing is performance, with a secondary focus on ease of adding other interaction features and programming the animation. There will probably be no more than 2000 points at once, with those small error bars on each one. Zooming in, out, and panning around need to be smooth. If the most recent SVG libraries are decent at this, then perhaps the ease of using d3 will outweigh the increased setup for KineticJS, etc. But if there is a huge performance advantage to using a canvas, especially for people with slower computers, then I would definitely prefer to go that way.
Example of app made by the NYTimes that uses SVG, but still animates acceptably smoothly:
http://www.nytimes.com/interactive/2012/05/17/business/dealbook/how-the-facebook-offering-compares.html . If I can get that performance and not have to write my own canvas drawing code, I would probably go for SVG.
I noticed that some users have used a hybrid of d3.js manipulation combined with canvas rendering. However, I can’t find much documentation about this online or get in contact with the OP of that post. If anyone has any experience doing this kind of DOM-to-Canvas (demo, code) implementation, I would like to hear from you as well. It seems to be a good hybrid of being able to manipulate data and having custom control over how to render it (and therefore performance), but I’m wondering if having to load everything into the DOM is still going to slow things down.
I know that there are some existing questions that are similar to this one, but none of them exactly ask the same thing. Thanks for your help.
Follow-up: the implementation I ended up using is at https://github.com/zooniverse/LightCurves
Fortunately, drawing 2000 circles is a pretty easy example to test. So here are four possible implementations, two each of Canvas and SVG:
These examples use D3’s zoom behavior to implement zooming and panning. Aside from whether the circles are rendered in Canvas or SVG, the other major distinction is whether you use geometric or semantic zooming.
Geometric zooming means you apply a single transform to the entire viewport: when you zoom in, circles become bigger. Semantic zooming in contrast means you apply transforms to each circle individually: when you zoom in, the circles remain the same size but they spread out. Planethunters.org currently uses semantic zooming, but it might be useful to consider other cases.
Geometric zooming simplifies the implementation: you apply a translate and scale once, and then all the circles are re-rendered. The SVG implementation is particularly simple, updating a single “transform” attribute. The performance of both geometric zooming examples feels more than adequate. For semantic zooming, you’ll notice that D3 is significantly faster than Protovis. This is because it’s doing a lot less work for each zoom event. (The Protovis version has to recalculate all attributes on all elements.) The Canvas-based semantic zooming is a bit more zippy than SVG, but SVG semantic zooming still feels responsive.
Yet there is no magic bullet for performance, and these four possible approaches don’t begin to cover the full space of possibilities. For example, you could combine geometric and semantic zooming, using the geometric approach for panning (updating the “transform” attribute) and only redrawing individual circles while zooming. You could probably even combine one or more of these techniques with CSS3 transforms to add some hardware acceleration (as in the hierarchical edge bundling example), although that can be tricky to implement and may introduce visual artifacts.
Still, my personal preference is to keep as much in SVG as possible, and use Canvas only for the “inner loop” when rendering is the bottleneck. SVG has so many conveniences for development—such as CSS, data-joins and the element inspector—that it is often premature optimization to start with Canvas. Combining Canvas with SVG, as in the Facebook IPO visualization you linked, is a flexible way to retain most of these conveniences while still eking out the best performance. I also used this technique in Cubism.js, where the special case of time-series visualization lends itself well to bitmap caching.
As these examples show, you can use D3 with Canvas, even though parts of D3 are SVG-specific. See also this force-directed graph and this collision detection example.