Python PackagingOct 28, 2012 by Andy R. Terrel
There are two major hurdles to Python disrupting the entire HPC work stack: packaging and dynamic loading. Today I want to discuss the packaging issue. While there are many people working on these hurdles, it is my opinion that the community needs to seek out methods to solve these hurdles together in a satisfactory way. I see this problem taking shape in many different communities, but the HPC version of the problem is probably the most difficult, thus by solving it, we can provide solutions for many other communities as well.
To this end, I helped host a conference call among the young, enthusiastic NumFOCUS community. This call seemed much more of a get to know the problem rather than a listing of solutions. We are working on editting the call and hope to have it published as a InSCIght podcast soon. The call included several companies that produces packaged Python solutions, university folks from around the globe, and industrial users of Python. The interest in the call was so great that we had to switch mediums at the last moment and lost out on some interactions with other great folks. I hope to have another call in the future and a discussion at the upcoming SuperComputing 2012 conference.
What is wrong with Python Packaging
Perhaps the place to start is defining the problem. When you have a piece of
code in a pure Python setting, one easily adds the top source directory to
PYTHONPATH and happily imports away. This model works quite well
usually. For the first years of Google App engine, code could only be installed
in this manner.
The issue is that in HPC one must depend upon libraries that are highly tuned for specifics of the architecture where the code is run. These highly tuned libraries often include assembly and usually only include compiled static or shared libraries. Since, HPC often uses specific compilers dedicated to their hardware, these libraries depend on the formats compatible with those compilers. To further enhance the problem, if these libraries are distributed via MPI, then the libraries depend on the MPI implementation (which depends upon the compiler). At TACC we have a hierarchical module system to keep this all straight. You load the compiler, then the MPI implementation, and finally any libraries. If you switch to a different compiler, the entire stack is reloaded. Actually a big part of my job is to make sure these stacks are all compiled in a way that they can be switched out, something that is a feat with biology codes. In a nutshell, we have a huge dependency chain that is binary and machine dependent.
Python isn't supporting this long list of dependencies very well. For HPCers this is not new as we are use to having to hack build and package systems. In fact in the eight years I've been around the FEniCS community, I've seen four package systems come and go. It took me three weeks to get the Trilinos code installed on my system back in 2006, at the time folks opined that picking that the right Trilinos configure was NP-Hard. It is much better today. In both cases though, CMake has turned out to be the tool that has stabilized everything.
Why can't Python support this jumble of dependencies? The reasons are many, but as David Cournapeau has pointed out, Python fundamentally mixes both build and packaging together in a way that makes it impossibly hard. I don't know David's history, but he seems pretty critical of the community about not understanding the issue. Rather than rehash all the arguments, I'm going to outline my view of the solution, separating build and packaging.
The build process
The difficult part of the build process is not calling the compiler, rather configuring the environment and compiler flags to produce the desired binary.
To give a simple example, if you are compiling a threaded library, say using OpenMP, you need to know that gcc uses -fopenmp, icc uses -openmp, pgi -mp, something else on xl, and clang doesn't even support it. Thank you Apple for making my job that much harder by using a default compiler that doesn't support OpenMP.
Okay that's actually not hard to code, but it gets better. Because multicore threading is still an infant technology, if you compose several threaded libraries you can see a real performance degradation (despite @dbeaz thinking threading is easy). Thus most vendors provide a threaded version of BLAS and a sequential version, the idea being if you have a big matrix to work with make BLAS be the multithreaded portion of the code, but if you have lots of smaller matrices let the application be multithreaded. This means that each application could be built with 4 compilers X 2 BLAS threading modes, we also need to throw in debugging and optimized modes as yes -O3 versus -g -O0 really do matter.
In this trivial example, we already need to test 16 different builds of a library, not to mention figuring out how many threads to run (which is often a runtime decision with OpenMP). If you want to drive your graduate students to madness get them to install Scalapack on your cluster.
At this point, you might say "16 versions of the library, quit your crying, baby!". It should be easy to script except Python distribution tools really don't give a way to manage these different compiler flags. In many cases the tools just slaps your code in some place on the path and runs all sorts of bootstrapping patching hacks to install at all. Virtualenv give a bit of sanity as one can isolate different Python builds, but it doesn't handle any of the system libraries at all. It is my understanding that when EPD was built, they initially tried to fix this problem but quickly switched to a single binary only distribution model.
To tackle all these burdens, we need to start using real configure and build tools. They need to work on all platforms, yes Windows matters in HPC. They need to be free and open source. And they actually already exist, the community just needs to start adopting them and that will take some effort. The two tools I've seen used well are Bento produced by David and CMake produced by Kitware. I'm not going to have a bake-off on the pros and cons of each, I have far more experience with CMake but people I trust tell me Bento is far easier to use. If NumFOCUS and the PSF really wants to help the community today, it would start helping people augment distutils to use one of these tools.
The package selection process
Now that we've discussed the build process, we need to package up those builds and allow developers to use them. I was really glad the Yaroslav, one of the Debian developers, and Samuel, homebrew's Python guy, was able to join us on the call. Open source platforms have created great packaging tools, but it only gets us code monkeys about halfway. I can install the basic libraries that aren't cutting edge and don't need fine grain configuration system wide. Then I just keep the jumbled mess in my developer space. But what if a developer needs to fix a bug with one of those system installed libraries that isn't the version installed? I've spend more time futzing with the packaging system installer than fixing the bug.
The truth of the matter is we really want the jumbled mess to be managed better too. Just like on my supercomputer, I can issue a single command and have a whole new compiler stack working, why can't I do that with my current development tree. Here enters HashDist. HashDist is Dag and Ondrej's plan for fixing this problem. The idea is to build libraries and stash them some place, recording the necessary details in a small distribution file that is hashed and put in a database store. Now when you need to build different versions of a library, one can simply refer to the different hashes of the other built libraries. At this point, HashDist is vaporware that has some funding. I have been working with these guys trying to fund this project for about a year and a half.
Let me be frank. If this tool existed, reporting bugs could come with a hashdist number that would set your machine up in the bug state immediately, i.e., no more futzing with installers just working. For HPC systems this would be huge. Hell it is even a good idea for a SAAS startup company.
This is very similar to the module system that we use at TACC, lmod, but it should build into the development cycle not just running a supercomputer. For what its worth, supercomputers use module systems because when you have 5000 users on a machine, you have to keep the kids from stepping on each others toes.
Why this affects developers outside of HPC
Hopefully, I've explained the problem well enough and pointed to a few promising fixes. While these problems are accentuated in HPC systems, they exist everywhere. Every single company, I have consulted with has this problem. They solve it different ways, but they all manage their own build and packaging system. Every single open source scientific Python team, has this problems. They usually solve it the old fashioned way, indentured servants, err graduate students. But perhaps more importantly, as PyPy and other python distributions threaten CPython for popularity, even Python core is having these problems. The wheel proposals are a good first step, but its gonna take more than changing package names to fix it.
NumFOCUS was founded to help encourage businesses and institutions to put funding together to solve science software problems. This is a big problem and one that prevents a large number of people from effectively using our ecosystem. I would like to see more funds and community resources put forward to create working solutions. Finally a homework for everyone who has read this far. I invite you to send me your versions of how you fix Python packaging for your work so I can collect the best ideas.