Speed Expectations for Julia

By: Simon Danisch

Re-posted from: http://randomfantasies.com/2015/05/speed-expectations-for-julia/

In this blog post I want to analyse the speed of Julia a little bit.
It is a very tedious task to write representative benchmarks for a programming language.
The only way out is to rely on a multitude of sources and try to find analytical arguments.
Julia’s own benchmark suite will be used in addition to two other benchmarks that I’ve found to be relevant.
In addition, the general compiler structure of Julia will be shortly analyzed to find indicators for Julia’s overall performance.
juliabench
(data from julialang)
In this first benchmark we can see that Julia stays well within the range of C Speed.
Actually, it even comes second to C-speed with no other language being that close.

Adding to this, Julia also offers a concise and high-level coding style. This unique combination of conciseness and speed is well illustrated in this graphic:

julia_codevstime

(different view of the data from julialang)

This is a very promising first look at Julia, but it should be noted, that these benchmarks are mainly written by the Julia core team.
So it is not guaranteed, that there is not an (unintentional) bias favoring Julia in these Benchmarks.

There is another benchmark comparing C++, Julia and F#, which was created by Palladium Consulting which should not have any interest in favoring one of the languages.
They compare the performance of C++, Julia and F# for an IBM/370 floating point to IEEE floating point conversion algorithm. This is part of a blog series written by Palladium Consulting.
F# comes out last with 748.275 ms, than Julia with 483.769 ms and finally C++ with 463.474 ms.
At the citation time, the Author had updated the C++ version to achieve 388.668 ms.
It does not say if the author put additional time into making the other versions faster as well, so it can not be said that the other versions could not have been made faster too.

The last Julia benchmark is more real world oriented.
It is comparing Finite Element solver, which is an often used algorithm in material research and therefore represents a relevant use case for Julia.

N Julia FEniCS(Python + C++) FreeFem++(C++)
121 0.99 0.67 0.01
2601 1.07 0.76 0.05
10201 1.37 1.00 0.23
40401 2.63 2.09 1.05
123201 6.29 5.88 4.03
251001 12.28 12.16 9.09

(taken from codeproject.)
These are remarkable results, considering that the author states it was not a big effort to achieve this. After all, the other libraries are established FEM solvers written in C++, which should not be easy to compete with.

This list could go on, but it is more constructive to find out Julia’s limits analytically.
Julia’s compilation model can be described as statically compiled at run-time. This means, as long as all types can be inferred at run-time, Julia will have in the most cases identical performance to C++. (See Julia’s performance tips, to get an idea of what needs to be done in order to achieve this)
The biggest remaining difference in this case will be the garbage collection.

Julia 0.3 has a mark and sweep garbage collector, while Julia 0.4 has an incremental garbage collector.
As seen in the benchmarks, it does not necessarily introduce big slowdowns.
But there are issues, where garbage collection introduces a significant slow down.
Analyzing this further is not in the scope of this blog post, though.
But it can be said that Julia’s garbage collector is very young and only the future will show how big the actual differences will be.

Another big difference is the difference in between different compiler technologies.
One of LLVM’s biggest alternatives is gcc, which offers similar functionality to LLVM’s C/C++ language front-end Clang.

If C++ code that is compiled with gcc is much faster than the same code compiled with LLVM-Clang, the gcc version will also be faster as a comparable Julia program.
In order to investigate the impact of this, one last benchmark series will be analyzed.
This is a summary of a series of articles posted on Phoronix, which bench marked gcc 4.92 against LLVM 3.5 and LLVM 3.5 against LLVM 3.6.

Statistic Speedup gcc 4.9 vs LLVM 3.5  Speedup LLVM 3.6 vs 3.5
mean 0.99 0.99
median 0.97 1.00
maximum 1.48 1.10
minimum 0.39 0.88

Bigger is better for LLVM.

Sources: gcc vs LLVMLLVM 3.5 vs LLVM 3.6

The results suggest, that LLVM is well in the range of gcc, even though that there can be big differences between the two.
These are promising results, especially if you consider that LLVM is much younger than gcc.
With big companies like Apple, Google, AMD, Nvidia and Microsoft being invested in LLVM, it is to be expected that LLVM will stay competitive.

So Julia should also stay competitive as it directly profits from any developments in LLVM.

This makes Julia a relatively save bet for the future!