R, OpenMP, MKL, Disaster
jyotirmoy.net
jyotirmoy.net
OpenBLAS is incompatible with application threads. Most Linux distributions provide a multi-threaded OpenBLAS that burns in a fire if you use it in multi-threaded applications. Even though OpenBLAS' performance is great, I'd be careful to give a general recommendation for people to rely on OpenBLAS. Like this MKL example, you have to be aware of its threading issues, read the documentation and compile it with the right flags (in a multi-threaded application: single-threaded, but with locking).
it's worth noting that OpenBLAS is as fast as MKL
This depends highly on the application. E.g. MKL provides batch GEMM, which is used by libraries like PyTorch. So if you use PyTorch for machine learning, performance is still much better with MKL. Of course, that is if you do not have an AMD CPU. If you have an AMD CPU, you have to override Intel CPU detection if you do not want abysmal performance:
https://danieldk.eu/Posts/2020-08-31-MKL-Zen.html
https://www.agner.org/optimize/blog/read.php?i=49
The BLAS/LAPACK ecosystem is a mess. I wish that Intel would just open source MKL and properly support AMD CPUs.
Can you explain what you mean by this? Are you saying there's a correctness issue here? I only recall running into issues with MPI, where you (typically) run one MPI rank (process) per CPU core. Then if you combine that with a multi-threaded BLAS library you'll suddenly have N^2 BLAS threads fighting over the CPU's and performance goes down the drain. The solution to this is, like you say, to use a single-threaded OpenBLAS, or then the OpenMP OpenBLAS and set OMP_NUM_THREADS=1
I guess with threads you'll have the same issue if you launch N cpu-bound threads and all those call BLAS, resulting in the same N^2 issue as you see with MPI.
There is a nice description of this:
https://github.com/xianyi/OpenBLAS/issues/2543
At a previous employer, we have seen various issues, including crashes, non-determinisms, etc. Usually, these issues would go away when switching to MKL.
That R code has since been ported to Python, but we faced the same issue again when using ThreadPoolExecutor, so we had to change it into ProcessPoolExecutor instead.
A year ago, I benchmarked a transformer network with libtorch linked against various BLAS libraries (numbers are in sentences per second, MKL with CPU detection override on AMD, 4 threads):
Ryzen 3700X - OpenBLAS: 83, BLIS: 69, AMD BLIS: 80, MKL: 119
Xeon Gold 6138 - OpenBLAS: 88, BLIS: 52, AMD BLIS: 59, MKL: 128
I guess people avoid AMD's support, because MKL is just much faster? AMD BLIS did add batch GEMM support since then. Didn't have time to try that out yet.
We don't know what that example was actually measuring, except apparently not the same thing for BLIS and MKL. On the basis of only that, it's not reasonable to say "just so much faster", in particular for what I care about. I have Zen2 measurements (unfortunately only in a VM) using the BLIS test/3 framework. MKL came out nearly as fast as vanilla BLIS 0.7 and OpenBLAS on serial DGEMM, less so on the rest of D level 3, and nowhere close with S, C, and Z. Similarly for one- and two-socket OpenMP. At least in that "2021" version of MKL, there's only a Zen DGEMM kernel.
Given that their latest compilers are based on LLVM, that seems like a fair trade between the closed- and open-source worlds.
I’ve never had any issue when using it in OpenMP codes (either compiling it myself or using the libopenblas_omp.so present in some distros), what do you mean by “burn in a fire”?
R is single-threaded.
Also in a previous life, I recall running into distro openblas packages that were not compiled with DYNAMIC_ARCH=1 (which enables the openblas runtime cpu target architecture selection, similar to e.g. MKL) but were instead compiled with some lowest common denominator x86_64 arch. I filed some bug(s?), and IIRC this problem has subsequently been fixed.
Huh? The bug tracker is here:
Yes, for filing a bug you need to request an account because they apparently were overwhelmed with spam, as documented here:
It turns out the way the second R package would determine the required precision of floats in sparse arrays was based on the compiled linear algebra libraries available. It took a week for us to debug and ultimately it was easier for us to just rewrite the whole thing in Python.
Renv has made things easier but I don't think packrat/renv allows you to lock C/C++ libraries as well as R ones.
gcc -shared -Wl,-soname=libgomp.so.1 -o libgomp.so.1 empty.c -lomp5
That is, instead of checking after doing the x[i] *= SCALE bit with cblas, I would check both before and after the scaling.
I want to add another dimension to this argument, what if we can maintain an existing language eco-system (library, community, etc) but modernize the engine that's running and compilation of the R language. This new engine can avoid the dreaded global locking limitation, provides native multi-thread applications and seamless interface with non-native R libraries in C/C++. Interestingly someone has tried this, with a sponsorship from Oracle no less, and presented this futile effort in the last year's R flagship conference keynote [2].
IMHO he will be more successful in his endeavour if using D language in his previous endeavors. What so special about D language you may ask? I would refer to the fact most of the languages do not provide Ruby on Rails (RoR) like tool except D but that for another story (see ref [1]). There's also the fact that D has a working alternative library to OpenBLAS and MKL, and it's even faster than both them five years back [3]! D also supports open method as an alternative for multiple dispatches that is much touted by Julia language community. D is also bringing native support for borrow checker feature that's always mentioned in the same sentence as Rust language. In addition D also has second to none FFI support for C and C++ language. Heck the latest D compiler has standard C compiler built-in. I can go on furthermore but I think you've probably already got the pictures.
My not so humble proposal to R and D language community is to compile R on top of D language. Essentially you a have dynamic language of R that is compiled at runtime (CTFE) on top of static D language. This approach is becoming more popular now as posted recently for the new Val and Valet language combination [4]. Just think of CTFE as the new JVM, but provides truly static and native compilation for R.
[1] What makes a programming language productive? “Stop designing languages. Write libraries instead.”:
https://jaxenter.com/stop-designing-languages-write-librarie...
[2] Why R? 2020 Keynote - Jan Vitek - How I Learned to Love Failing at Compiling R:
https://www.youtube.com/watch?v=VdD0nHbcyk4
[3] Numeric age for D: Mir GLAS is faster than OpenBLAS and Eigen:
http://blog.mir.dlang.io/glas/benchmark/openblas/2016/09/23/...
[4] Show HN: Val - A powerful static and dynamic programming language(val-lang.org):
R is good for explorative data analysis but useless for everything else.
I hear a lot of "R is bad, Python is Enterprise Production Quality (TM)" blather at my work. It's always because the people involved don't understand computers, don't read documentation, don't debug, don't do root cause analysis, and want to quickly pass off responsibility for their laziness and incompetence. Meanwhile I and my team are happily chugging away, producing millions of dollars of reliable value for my company in R year after year.
Python lags far behind R in wide swaths of data science. Pandas is inferior to both dplyr and data.table, and R's modeling capabilities blow Python's out of the water in breadth and depth. You only use Python when you have to, e.g. for unstructured data and deep learning type stuff.
If your colleagues make you deal with their bad R code, that's too bad, but don't blame the language. It's designed to be easy to use, so a lot of bad coders use it. Go train your bad coders or hire better ones.
Same issue as Excel, really. Easy to use, so you get a lot of users with very thin engineering skills.
The solution is for production engineers to understand just enough R to set standards for data scientist code that enable reliable translation of the models to the production language. As with JS, you can complain about the yucky parts, or you can accept that it's the best tool for some jobs and make an effort to work around the yucky parts, or use the tools of those who are doing that (e.g. tidyverse and Wickham).
If you want data scientists to produce production-ready results, you have to hold them to the standards of production engineering.
Huh?
While I totally agree with your quote, I'd think it applied a lot more to python than to R. Especially given that python seems to be the dominant "first language for people to learn when they get into programming" because it is "easy".
Python isn't much better in this regard, thanks to the GIL.
What I actually found most baffling when I delved into R is the fact that it doesn't support 64bit integers (lack of proper native UTF-8 support coming a close second).
Take some standard ML model built with Caret or LME4 and try serve predictions with Plumber in R. It’s significantly more painful than using sklearn + FastAPI. You either need to use future::promise (which still sucks because it’s forking new R runtimes) or forgo this and go K8s or something similar.
I don’t get the love for RStudio either. It crashes frequently for me, or locks up randomly. The debugging experience is abysmal compared to PyCharm. Getting reproducible R builds are a pain, slightly alleviated by Renv. But not really if you want separate dependencies for dev and production.
Python and R tooling are not comparable. You will have serious issues operationalising R. Skills that most statisticians are simply not equipped to deal with, and serious software engineers will hate about R.
Unintentional and unnecessary creation of huge, memory-hogging objects is a closely related footgun. Packages are often not built with large data in mind and make choices that scale terribly, such as storing multiple copies of the data in the model object, or creating enormous nonsparse matrices to represent the model term structure. It's a legacy of the academic statistics culture R grew out of. Most researchers test their fancy new method on a tiny dataset, write a paper, and call it a day.
No argument about the debugging experience. I find it very slow, especially with large datasets, and try to avoid it. Not much experience with reproducible R builds but I wouldn't be surprised if it was a pain.
Now some people say this can be solved with a good IDE. Which might (or might not) be true if you can reliably identify, by manually reviewing the code, the ends of the functions, loops, etc which got munged in the paste.
But interestingly enough, jupyter notebooks (which seem to be the go-to tool these days) aren't IDEs. Making it incredibly easy to fubar otherwise perfectly working code by pasting it from your local IDE into, let's say an AWS Sagemaker instance, to pick one random example of a current widely used jupyter implementation. So even if the problem could be fixed by a good IDE, there is no guarantee that that IDE is (easily) accessible for production code.
I just have a hard time seeing how such a fundamental flaw in a language can lead to "good software engineering"
A language that's more flexible than your favorite "encourages bad habits", while a language that's less flexible than yours is "bureaucratic".
R encourages bad habits for the following reasons:
- R is made "for statisticians, by statisticians" so a lot of the example code out there is very poorly written
- The syntax is very inconsistent across libraries, and even within base R
- There are a lot of syntactical quirks that cause a lot of confusion for anyone who's learned another language, like using dots in function tables, e.g. "read.csv". There's also the 1 indexing.
People often say "LISP is so powerful but nobody can understand anyone else's code". That's the dominant explanation for why LISP isn't more popular (along with the "oatmeal with toenail clippings mixed in" syntax, which most people don't find readable, regardless of the fervent beliefs of the LISP community to the contrary). The community stays small because of the unappealing syntax, and even within the community people find it hard to work together, because everyone has their own style, so the kind of coding and collaboration that produces generally useful libraries doesn't tend to happen. I would argue there's no meaningful distinction between "bad habits" and "habits that inhibit the development of generally useful software". In fact, that's the most useful definition of "bad habits" I can imagine.
Flexibility is a root cause of bad habits thus defined, because flexibility is what enables people to make bad choices. Language designers have long recognized this. It's why certain languages impose heavy restrictions on how you can structure your code, from Java's everything-is-a-class to Python's indentation-based scoping. They have a certain vision for what constitutes effective code and they know it won't happen unless they force everyone to follow it. In other words, they choose to reduce flexibility to prevent what they consider to be bad habits.
> R encourages bad habits for the following reasons:
Your reasons are just common complaints about R issues, not actual arguments that these issues encourage bad habits.
> - R is made "for statisticians, by statisticians" so a lot of the example code out there is very poorly written
At best this is an argument that widely publicized badly written R code encourages bad habits, not R itself.
> There's also the 1 indexing
There's a big difference between "language feature I don't like" and "language feature that encourages bad habits". A language that has different conventions than your favorite language is just different.