Back in 2012 a paper by four computer scientists including Alan Edelman of MIT introduced Julia, A Fast Dynamic Language for Technical Computing. At the time, the gold standard programming languages for fast performance on computationally intensive problems were C and Fortran – maybe they still are. Fast forward to 2020 and Julia is making a run at the top and Edelman won last year’s IEEE Sidney Fernbach Award presented at SC19 for, among other things, his work on Julia.
Consider recent stats on Julia adoption. By January 1, 2019, reports Julialang.org, the total downloads of Julia reached 7.3 million. That number jumped to 12.9 million, a 77 percent increase, by January 1, 2020. The number of published citations for same period rose 66 percent from 1048 to 1680. In his SC19 talk, Edelman noted that as of October 2019 there were 3,119 Julia packages available, up from 1,688 at the year’s start. Those numbers are impressive all around.
In September, Julia joined the ranks of computing languages that have achieved peak performance exceeding one petaflop per second – the so-called ‘Petaflop Club.’ The Julia application that achieved this milestone is called Celeste. The Celeste team developed a new parallel computing method to process the entire Sloan Digital Sky Survey dataset and loaded an aggregate of 178 terabytes of image data to produce the most accurate catalog of 188 million astronomical objects in just 14.6 minutes. Celeste achieved peak performance of 1.54 petaflops using 1.3 million threads on 9,300 Knights Landing (KNL) nodes of the Cori supercomputer at NERSC – a performance improvement of 1,000x in single-threaded execution.
You get the idea. Julia is on a roll. Whether Julia will challenge Python the way Python once challenged and then surpassed Java is an interesting question being bandied about. What’s clear is that after percolating along steadily during its early years, Julia use is growing quickly – much to Edelman’s delight.
The tension between the high performance delivered by so-called static programming languages and the lesser performance delivered by high-level programming languages, which emphasize abstraction, speed of development, and portability, hasn’t gone away, noted Edelman. But convenience with sufficient performance is winning out. Moreover, the rise of heterogeneous computing and the complications it presents to programmers has increased the tilt away from static programming.
In his engaging SC19 talk, Edelman noted:
“When you’re writing various algorithms, you don’t necessarily want to think about whether you’re on a GPU, or whether you’re on a distributed computer. You don’t necessarily want to think about how you’ve implemented the specific data structure. What you want to do is talk about what you want to compute, not how you want to compute it, right? That is the big problem, to get people to talk about what you want to compute, and not how you want to compute it. Because if you put in your software, how you’re going to compute it, and if your software is filled with that muck, I promise you, nobody’s ever going to change it. No one’s going to innovate on it. When the person who wrote it is no longer in the project, no one’s ever going to touch it.
“[S]ome of the reasons why Julia is working very well is because we have particularly well-designed abstractions. We have something called multiple dispatch and we have a very careful balance between the static and dynamic. It interfaces with LLVM. It plays nicely with Python. We also have had lots of people take legacy codes in MPI, and plug them into Julia – you don’t get all the benefits, but what you do have, which might be the most important benefit, is other people can now run your code once it’s inside of Julia. So it’s much easier for other people. You can actually give your old code new life when you plug it into a higher level language.”
Julia, say advocates, minimizes performance penalties because it was designed from the outset with parallel computing in mind and with making use of high-performing abstractions able to exploit the latest libraries and deliver portability. Edelman presented an example in which a group of researchers decided to scrap their legacy climate code in Fortran and write it from scratch in Julia. There was some discussion around performance tradeoffs they might encounter in the move to a high level programming language. The group was willing to accept a 3x slowdown for the flexibility of the language. Instead, said Edelman, the switch produced 3x speedup.
He briefly presented a second example in which Julia was used with GPUs and skirted CUDA.
“I’m going to go over really fast a little bit on how we do Julia on GPUs. Because we have these different levels of abstraction, we’re able to reason about what’s going on at various different levels. If you only have that very lowest level on a GPU again, you can roll up your sleeves and work really hard, but you don’t get any code reuse. We have the saying in the Julia world, where if you’ve copied and pasted code, and you just modified a few things, then you’ve done something wrong. We’re trying to eliminate the copying and pasting the code, not only by you, but somebody else in the community shouldn’t have to go and take somebody else’s code and copy and paste it. That shouldn’t be necessary. So there’s a lot going on in here (see slide below). But the main point is that it’s not just queued in Julia, but it’s actually Julia running on the GPUs. And so that’s, that’s pretty exciting.”
Edelman’s perspective on coding and mathematics is interesting. “You know, a lot of us get this impression from universities and from teaching that you learn some math and then you [build] an algorithm, and then you code it up as if there is the algorithm first, and then the coding is sort of the secondary. But you know, more and more now, the code is the math,” said Edelman. This idea, he suggests, should inform our thinking about coding generally; it’s yet another effective abstraction.
The video is best watched (or listened to) to get the breezy yet substantive flavor of Edelman’s ideas and Julia’s capabilities.
“Julia was always designed to be a high level of parallel computing language, even from day one. That’s what I wanted. This is the problem that I personally wanted to see solved. We’re not fully there yet. But Julia is a highlight. You could do distributed computing, you could do GPU computing, you could do shared memory computing. We have models, you know, asynchronous computing, whatever you’d like to do we have models to do it now. And the real question, the one that everybody asked, the one that none of us really knows how to do, the deep intellectual problem is how to put it all together. But if we all work together at it, if we actually all share code, and, you know, hammer away at it, I think we could actually solve this problem.”
Link to intro paper: https://arxiv.org/pdf/1209.5145.pdf
Link to Fernbach Award announcement: https://www.hpcwire.com/off-the-wire/julia-computing-chief-scientist-alan-edelman-wins-prestigious-ieee-sidney-fernbach-award/
Link to SC19 video: https://www.youtube.com/watch?v=nwdGsz4rc3Q