Prettify my Emacs symbols

Emacs Redux made me aware of a new nifty feature of Emacs 24.4 (the upcoming release; I am currently using one of the nightlies). Emacs 24.4 includes a new mode called prettify-symbols-mode, which replaces tokens within code with more compact symbols (typically unicode characters like λ).

I am using the mode to replace the “function” token in JavaScript — which is used all over the place, since it is the syntax to create functions, modules and closures — with the single character λ. This results in significantly more lightweight expressions:

    
initialize: λ() {
    this.listenTo(this, "planet:drag planet:dragvelocity", 
        λ() { this.elements(true); });
}

The mode is smart enough to only replace tokens (it doesn’t replace function within a string, e.g., var a = "This is a function"; will appear unmodified). To activate the mode, use (global-prettify-symbols-mode 1) in one of your init files, and add new symbols to the mode hooks:

(add-hook 'js2-mode-hook
            (lambda ()
              (push '("function" . ?λ) prettify-symbols-alist)))

Ever been annoyed with the Emacs bell? Aside from disabling it, you can make it slightly less annoying by using a visual error indicator in place of an audible bell. There’s a Visible Bell setting included by default, but I find it quite aesthetically displeasing (at least on my nightly Mac build). I like this snipped a lot better (from EmacsWiki):

 (defun my-terminal-visible-bell ()
   "A friendlier visual bell effect."
   (invert-face 'mode-line)
   (run-with-timer 0.1 nil 'invert-face 'mode-line))
 
 (setq visible-bell nil
       ring-bell-function 'my-terminal-visible-bell)

This snippet will flash your mode line instead.

Please excuse the mess.

Over the past week or so, I’ve been browsing WordPress themes over the web. Now, I’ve never been completely satisfied with WordPress. While its dashboard is great when you need to get a blog started, I’ve been finding it increasingly stifling — themes are opaque, the post editor is limiting (needs more Emacs!), and I have a general distaste for the large soup of tags generated for each page.

Today I foolishly tried a new theme, thinking that I could easily revert to my highly mangled Twenty-Fourteen theme at any time. Alas, it appears all the customizations were gone when I tried to do exactly that. Luckily, the new theme (Wilson by Anders Noren) is quite to my liking, and I only had to tweak a few things. I also learnt to create child themes in order to avoid losing my settings in the future.

Still, I feel like it is high time for me to move away from WordPress and its pre-made themes, to a simple static website with my own versioned assets. Jekyll looks like the clear winner — it would be well integrated into my current workflow of simple static HTML and CSS, and it would be easily uploaded on GitHub (automatic backup FTW!). Although I don’t have that many pages up, the idea of migrating all the pages and the posts is slightly daunting, so the migration of the blog will probably occur over a few weeks.

In the meantime, any brokenness on the website is due to the theme switching and will be resolved shortly :)

Data Science at the Command Line webcast

Yesterday, I attended a very handy webcast by Jeroen Janssens called Data Science at the Command Line (a book is on its way). While I do most of my data manipulation from R, it is undeniably convenient to be able to run some simple tasks interactively from the command line, or as part of a shell script or Makefile.

The presentation touched on several command tools that I either wasn’t aware of, or had forgotten about.  If you missed the webcast, this website has a helpful list of commands — of both the well-known and the obscure variety. Below are a few that I had not known about and might be useful to others.

parallel

parallel is a shell command to execute a series of commands in batch, over several CPUs on a local machine or over several computers (using a combination of ssh and rsync to connect and transfer files). I used to use a combination of Xgrid and some homegrown shell scripts to achieve that, but Xgrid is unfortunately no more (RIP!). parallel seems like a way  to quickly get a small subset of Xgrid’s functionality.

cowsay

cowsay is like echo, but with cute ASCII art animals.

~$ cowsay "I'd rather get error messages from a cute animal."
 ______________________________________ 
/ I'd rather get error messages from a \
\ cute animal.                         /
 -------------------------------------- 
        \   ^__^
         \  (oo)\_______
            (__)\       )\/\
                ||----w |
                ||     ||

Very useful to add some levity to your error messages! Use cowsay -l to get a list of animal templates.

I used the Node.js cowsay package to create this little Node.js app (for fun, and to learn a minimal amount of Express.js). Install via Homebrew or MacPorts.

asciinema

I saw Jeroen use this tool during the webcast to record his terminal session. Asciinema looks very very cool and potentially helpful to create terminal-based tutorials. Install via pip.

csvkit & jq

csvkit is a suite of command line tools to deal with CSV files, but works quite well for tab-separated data as well (which I deal with often). Particularly useful so far: csvlook (nicely formatted table in the terminal), csvstat (column statistics) and csvsql (SQL queries on CSV files). jq can be used to manipulate JSON data, potentially piped into csvkit to create simple text tables.

Hot fix for Systemic 2 — 2.171

I just posted a revision of Systemic 2 that corrects a small palette issue that arose during the packaging of 2.17. It’s up now!

Systemic wins LIFT award, new version of Systemic (2.17)

I am very happy to announce that Systemic has been awarded the LIFT grant. This means that, together with my colleagues Joel Green and Randi Ludwig, I will be able to work full-time — instead of as a side project — on improving and expanding Systemic, and creating new educational apps like Systemic Live and Super Planet Crash, over the next year!

I am also releasing a new release of Systemic 2 (2.17) which addresses some bugs and improves the documentation for installation on Linux. You can download it now.

Below are some of the changes:

- NEW: ktable function for listing the fit values as a table, suitable for exporting to TeX or HTML
- NEW: Bayesian Information Criterion menu item
- NEW: Quadratic trend term
- CHANGED: Periodograms report the normalized power between 0 < p < 1, where power = 1 is a perfect fit. — This is a bug corrected in 2.172.
- FIXED: bug where the semi-amplitude would be calculated incorrectly for massive bodies. (credit: Trifon Trifonov)
- FIXED: bug in simplex that would crash the application if the minimizer encountered a NaN value.
- FIXED: bug in the GUI that would crash the application in case of excessive text output.
- FIXED: Fixed naming of columns in the matrix returned by kdata().
- FIXED: bug where the radial velocity curve or periodogram would look excessively jagged.
- FIXED: bug in kperiodogram.boot where the function could crash.
- FIXED: bug in kperiodogram.boot where the function would only calculate the ‘full’ periodogram (instead of the periodogram of residuals) for certain inputs.
- FIXED: bug in the GUI periodogram routine, where you could receive an error for certain power spectra.
- FIXED: Kernel plot using plot() respects the chosen xlim.
- FIXED: MCMC would crash in certain situations when set up from the menu.
- FIXED: 1LO-crossval returns the signed sum of logs, instead of the absolute value.
- FIXED: clarified the installation instructions (Readme.txt) for Linux. (credit: Franz Feldtkeller)
- FIXED: you can now choose a path for R that is not /usr/bin/R by selecting Help -> Set path to R…
- FIXED: F-test menu item uses the current kernel instead of the kernel named “k”.
- Various bug-fixes in the plotting routines.

Bugs reported at https://github.com/stefano-meschiari/Systemic2/issues/2 are unfortunately still open due to lack of time to address them. They will be fixed in 2.18.

Colliding N-body spheres: Particle Mayhem!

When Giants Collide — WGC for short — is one of the “fun” projects I am working on. Once finished, it will be a small in-browser simulator where you can collide giant planets together (with some degree of realism). You can see my progress on my GitHub repo and the series of blog posts under this category.

In this little demo app, you can run an N-body simulation in your browser where you make two spheres made of point masses “collide”. You can tweak various parameters (collision speed,  impact parameter, distance and number of particles) to change the outcome of the simulation.


Underneath it all: TreeSPH.js

The app above is powered by the portion of WGC’s code that computes the gravitational force between a set of point particles (a gravitational N-body system). The gravitational force is computed using the Barnes-Hut tree gravity algorithm, and the coordinates of the points are evolved using a third-order embedded Runge-Kutta algorithm. The code is available in the GitHub repo for WGC.

I am now working on writing the hydrodynamical part (via the SPH algorithm), which will let me simulate the collision between two gaseous spheres. The resulting library will be called treesph.js.

treesph.js will be an open-source JavaScript library able to power small-scale hydrodynamical simulations — either in the browser (through web workers), or within a JavaScript environment (e.g. Node.js). It comes with:

- a library to set up initial equilibrium conditions (e.g. Lane-Emden spheres, or N-body spheres with isotropic velocity dispersion);
- a canvas-based library to plot and animate the simulation snapshots;
- a fast library for operating on vectors and matrices that minimizes allocations and copying, and other math routines.


Performance notes

While playing with the app, you may be wondering (a) why there is a “buffering” stage before you can see the evolution of the system, and (b) why so few particles?

Buffering: it’s all about delayed gratification

The app animates the particle motion at 30 frames per second. This requirement places a hard and fast constraint: if you want to compute the particle motion at each frame request, then the computation must take less than 1/30th of a second, otherwise the webpage will freeze as the JavaScript engine tries to catch up with the accumulated frame requests. For reasonable number of particles (say, 100 or more — see below) and the time steps required by the above app, this requirement is way overshot.

This issue can be ameliorated by running the numerical computation in a separate thread (a web worker), and drawing frames on the main thread as soon as they are computed. This is still not as optimal: while it solves the UI freezing issue, the particle motion will appear very jerky as it will be animated at (typically) less than a frame per second!

In order to solve this issue, I created a small JavaScript library (streamingcontroller.js; available in the same GitHub repo, documentation upcoming). Streamingcontroller.js first estimates the expected wall time — the time in seconds — needed to complete the simulation. Then, within the web worker thread, it “buffers” the simulation snapshots by adding them to a pool of snapshots. Once the buffer is big enough that the simulation can be run in real time without hiccups, the library starts streaming the snapshots back to the main thread where the animation is drawn. In the main thread, a second buffer receives the snapshots; the second buffer is then emptied at 30 frames per second.

More particles, pretty please?

The default setting of the app is to animate 250 particles (125 per sphere). Why so few, when typical number of particles quoted for N-body simulations routinely exceed millions — or even billions! — of particles?

There are three bottlenecks at work. The first is obvious: the code isn’t fully optimized and profiled yet, and I am certain there is room for improvement. I am writing a small math library of common mathematical routines called math.js (also in the same GitHub repo) which will be fully optimized for V8.

The second is also obvious: simulations with lots of particles are usually run at full-speed, on multiple cores, and in the background. These simulations can save their snapshots, to be plotted and animated at the end of the run. An online app (or game) with real-time requirements (or, say, a <1 minute buffering time) doesn’t have this kind of luxury!

The third is the worst hurdle, and it is inherent to the nature of JavaScript: JavaScript is slow. I am not a JavaScript guru by any means, but I do have a good amount of experience writing performant numerical code in a variety of languages (mostly C). While JavaScript is typically fast enough for most tasks on the web, it is slow on personal computers and even slower on mobile platforms for physically-motivated, accurate simulations. In its present form, it is not well-suited to run these kinds of numerical tasks as quickly as the underlying hardware allows. Although JavaScript interpreters have been improving by leaps and bounds, and careful code can exploit some of these optimizations, they are hitting a wall of diminishing returns. Since JavaScript is the only runtime available on browsers, it is the ultimate bottleneck.


You can check out the other demos using code from WGC in this webpage.

A Barnes-Hut tree gravity benchmark

In one of my last posts (An interactive Barnes-Hut tree) I talked briefly about one of the “fun” projects I’m working on, When Giants Collide (work in progress, GitHub repo), and promised myself to blog about its development as I went along. I just finished refining the algorithm for building the tree and calculating the gravitational force.

The small app above is a benchmark pitting the Barnes-Hut algorithm for computing gravity (an O(N log(N)) algorithm) against a brute-force direct summation (an O(N^2) algorithm). It calculates the gravitational field of a random collection of particles using both methods for N = 256 to N = 16,384; a lower amount of time spent indicates a faster algorithm. The time used to compute the gravitational force is averaged over 12 iterations to minimize fluctuations. Results are plotted in real time.

Lastly, it calculates an overall “score” for the JavaScript interpreter by only running the Barnes-Hut algorithm for N = 16,384. You can see a table of scores for a few different browsers and devices I have access to (lower is better). If you’d like, send me your score!

Some observations about JavaScript optimization

Chrome turned out to be the fastest browser at this particular benchmark. Surprisingly, a previous version of the same code was actually the  slowest on my MacBook — almost 6x as slow as Safari! That was quite unexpected, as in my (limited) experience building web apps Chrome tends to edge out other browsers in terms of JavaScript execution speed.

So I waded a little bit more into my code to understand what was making my code so inefficient. This Google optimization guide and this post on HTML5Rocks (specifically talking about optimizing for V8, the just-in-time compiler embedded in Chrome) proved very useful. What I learned:

  1. Use the idiomatic JavaScript style for creating classes (using prototypes, new, straightforward constructors etc.) instead of using an object factory and closures.
  2. Avoid creating closures, when possible.
  3. Use node.js to profile the application and identify functions that are not getting optimized (using –trace-opt).
  4. Both Safari and Firefox had good baseline scores even before these optimizations. I found it quite surprising that V8 was much more fastidious about my code than the other JavaScript engines.

Another finding was how much slower alternative browsers (e.g. Chrome, Mercury) are on iOS. Alternative browsers use the same engine as Safari, but they don’t have access to Nitro’s Just-In-Time compilation – this means that they will be quite a bit slower than Safari on a computationally-intensive benchmark. How much slower? On my iPhone 5S, almost a factor of 10!

Web workers are awesome

The benchmark runs in a different thread, so that the page itself remains responsive. This is accomplished using Web Workers, a relatively new technology that allows the page to spin off threads to do computation-heavy work. It’s quite well supported, and I found it pretty easy to learn (aside from some surprising quirks). I plan on spinning off some of the tasks in Systemic Live – which currently either block the interface or use timers — into Web Workers (it’ll be a quite a bit of work, so don’t hold your breath).

AstroBetter: Creating Online Apps for Outreach and Education

I wrote a short article on AstroBetter called “Creating Online Apps for Outreach and Education” about some of the tools and resources I use for my online work.

Chrome developer tools + js2-mode and web-mode in EmacsChrome developer tools + js2-mode and web-mode in Emacs

I’ll be blogging more about this topic in the short term, but here are a few things I’ve been looking at that people might find interesting:

  • Languages that convert to JavaScript: I find the constant context switching between the different programming languages I use for my work very annoying. I’ve used Emscripten for converting my C code into JavaScript to minimize the amount of code rewriting. I’m actively looking at languages that compile to JavaScript — the most famous is probably CoffeeScript, but I’m keeping an eye out for Kotlin. Kotlin is a budding language that primarily runs on the Java Virtual Machine (JVM), but can also be compiled to JS. Since I do use Java and the JVM ecosystem for a substantial amount of my code, it looks very promising. Now, if only there was a kotlin-mode for Emacs I’d be very happy :)
  • Three-dimensional visualizations: three.js looks cool but daunting. I’ve been trying to think of a small project to do using the isometric engine Obelisk.js – I’m just enamoured with things that look retro.

Cogito.org, a magazine for young students interested in science, published a short interview with me about how I came to be an astronomer, and why I developed Super Planet Crash. Read it here!

An interactive Barnes-Hut tree

[TL,DR: if you'd like to play with a simple Barnes-Hut octree code, scroll down to the little embedded app.]

Gah! It’s been quite a while since my last post. Despite my best intentions, work (and a lot of feedback from Super Planet Crash!) has taken precedence over blogging. I do have a sizable list of interesting topics that I’ve been meaning to write about, however, so over the next few weeks I’ll try to keep to a more steady posting clip.

Super Planet Crash has been a resounding success. I have been absolutely, positively astounded with the great feedback I received. My colleagues and I have been coming up with lots of ideas for improving the educational value of SPC, add new, interesting physics, and addressing some of the complaints. In order to have the ability to dedicate more time to it, over the past few months, we’ve been furiously applying for educational and scientific grants to fund development. Hopefully something will work out — my goal is to make it into a complete suite of edu-tainment applications.

When Giants Collide

I’ve recently started experimenting with a new  visualization that I think will turn out pretty darn cool. Its draft name is When Giants Collide. When  Giants Collide will address a common request from planetary crashers: “Can I see what happens when two giant planets collide”?

A sketch of the interface.

When Giants Collide will be a super-simple JavaScript app (so it will run in your browser) that will simulate the collision of two massive spheres of gas. The simulation will have to model both gravity and the dynamics of the gas: to address this, I’ve been dusting off and reviewing an old Smoothed-Particle Hydrodynamics (SPH) code I worked on for a brief period in graduate school. SPH is a very simple technique for cheaply simulating gas flows with good spatial accuracy, and is somewhat straightforward to code. There are some shortcuts that have to be taken, too — large time steps, low particle counts, and more (e.g., a polytropic equation of state for the gas giants; more on this in future posts). These shortcuts come at the expense of realism, but will enable fast, smooth animation in the browser.

Gravity with the  Barnes-Hut algorithm

Gravity is an essential ingredient of When Giants Collide! Even with very low particle counts (say, N = 1000), a brute force calculation that just sums up the mutual gravitational force between particles won’t do if you want to run the simulation at 60 frames per second. Direct summing is an N^2 operation:

(this is a simple force accumulator written in R).

A better way that involves only a slightly more complicated algorithm is to use the Barnes-Hut algorithm (a short Nature paper with more than 1,000 citations!). The algorithm involves recursively subdividing space into cubes and loading them with particles, such that every cube contains either 0 or 1 particles. This is represented in code with an oct-tree structure.  Once such a tree is constructed, one can calculate the gravitational force on a given particle in the brute-force way for close particles, and in an approximate way for distant particles; whether to use one or the other is determined by walking the tree down from the top. An excellent explanation (with great visuals!) is provided in this article.

The other advantage is that, once the tree has been already built for the gravity calculation, it can be used to identify the nearest neighbors of a given particle through the same tree-walking procedure. The nearest neighbors are needed for the hydrodynamical part of the SPH algorithm (see, e.g., this review article by Stefan Rosswog or this one by Daniel Price).

An interactive tree

Below is an interactive JavaScript applet that subdivides space with the Barnes-Hut algorithm. You can add new points by clicking on the surface, or using the buttons to add new, random ones.

The code for building the Barnes-Hut tree from an array of 3D positions is available at the GitHub repository for When Giants Collide. I will be developing the code in the open, and post periodically about my progress. Hopefully by the end of summer I will have an attractive app running on any modern device and web browser. Any ideas on how to gamify it?

The Automated Planet Finder, Systemic and Super Planet Crash

[This short article I wrote has been published on The Conversation UK.]

The following is a short article about the Automated Planet Finder, Systemic and Super Planet Crash. We recently announced the first batch of exoplanets that were discovered in the first few months of science operation of APF. The first two systems (HD141399 and Gliese 687) have been submitted and will be available on astro-ph shortly.


Telescope apps help amateurs hunt for exoplanets


Laurie Hatch

People around the world are being invited to learn how to hunt for planets, using two new online apps devised by scientists at the University of Texas at Austin and UC Santa Cruz.

The apps use data from the Automated Planet Finder (APF), Lick Observatory’s newest telescope. The APF is one of the first robotically operated telescopes monitoring stars throughout the entire sky. It is optimised for the detection of planets orbiting nearby stars – the so-called exoplanets.

Systemic is an app that collects observations from APF and other observatories and makes them available to the general public. Anyone can access a simplified interface and follow the steps that astronomers take to tease a planetary signal out of the tiny Doppler shifts collected by the telescope.

Students and amateurs can learn about the process of scientific discovery from their own web browsers, and even conduct their own analysis of the data to validate planet discoveries.

The second app, SuperPlanetCrash, is a simple but addictive game that animates the orbits of planetary systems as a “digital orrery”. Users can play for points and create their own planetary systems, which often end up teetering towards instabilities that eject planets away from their parent stars.

First catch

Despite only being in operation for a few months, APF has already been used to discover new planetary systems.

Night after night, the telescope autonomously selects a list of interesting target stars, based on their position in the sky and observing conditions. The telescope collects light from each target star. The light is then split into a rainbow of colours, called a spectrum. Superimposed on the spectrum is a pattern of dark features, called absorption lines, which is unique to the chemical makeup of the star.

When a planet orbits one of the target stars, its gravitational pull on the star causes the absorption lines to shift back and forth. Astronomers can then interpret the amplitude and periodicity of these shifts to indirectly work out the orbit and the mass of each planet.

This method of detecting exoplanets is dubbed the Doppler (or Radial Velocity) technique, named after the physical effect causing the shift of the absorption lines. The Doppler technique has been extremely productive over the past two decades, leading to the discovery of more than 400 planet candidates orbiting nearby stars – including the first exoplanet orbiting a star similar to our own Sun, 51 Pegasi. To conclusively detect a planetary candidate, each star has to be observed for long stretches of time (months to years) in order to rule out other possible explanations.

The APF has now found two new planetary systems surrounding the stars HD141399 and Gliese 687.

HD141399 hosts four giant, gaseous planets of comparable size to Jupiter. The orbits of the innermost three giant planets are dramatically more compact than the giant planets in our Solar System (Jupiter, Saturn, Uranus and Neptune).

Gliese 687 is a small, red star hosting a Neptune-mass planet orbiting very close to the star: it only takes about 40 days for the planet to complete a full revolution around the star.

Team leader Steve Vogt of the University of California, Santa Cruz has dubbed both of these almost “garden variety” planetary systems, and indeed, they are quite similar to some of the systems discovered over the last few years. However, what look like distinctly unglamorous planetary systems now can still pose a puzzle to scientists.

The new normal

The planetary systems discovered so far are typically very different from our own solar system. More than half of the nearby stars are thought to be accompanied by Neptune-mass or smaller planets, many orbiting closer than Mercury is to the Sun. In our solar system, on the other hand, there is a very clear demarcation between small, rocky planets close to the Sun (from Mercury to Mars) and giant planets far from the Sun (from Jupiter to Neptune). This perhaps suggests that planetary systems like the one we live in are an uncommon outcome of the process of planet formation.

Only further discoveries can clarify whether planetary systems architected like our own are as uncommon as they appear to be. These observations will need to span many years of careful collection of Doppler shifts. Since the APF facility is primarily dedicated to Doppler observations, it is expected to make key contributions to exoplanetary science.

The two apps produced by the APF team make amateur scientists part of the hunt. These applications join the nascent movement of “citizen science”, which enable the general public to understand and even contribute to scientific research, either by lending a hand in analyzing massive sets of scientific data or by flagging interesting datasets that warrant further collection of data.

The Conversation

© 2014 Synthetic Worlds

Theme by Anders NorenUp ↑