One of the goals of the OpenDreamKit project is to improve support for open source mathematics software on a wider range of hardware platforms and operating systems (see Task 3.1: Portability). Among the largest portability challenges is improving installation and operation of such softwares on Microsoft Windows–still the dominant OS in many user communities, especially on desktop and laptop computers. Despite there being many large communities of Windows users, most open source software developers have traditionally preferred UNIX-like software development environments. The UNIX environment differs in many significant ways from Windows, such that support for Windows has often been neglected by those developers.
As part of our series reviewing emerging technologies, we look at a new development from Microsoft that has the potential to open up a smörgåsbord of open source mathematical and scientific software to Windows users that was never previously available to them.
Introducing Windows Subsystem for Linux
In late March of 2016, at its annual developers’ conference, Microsoft announced a surprising new technology. Dubbed Windows Subsystem for Linux (WSL), this new feature premiering in the Windows 10 “Anniversary Update” would add a Linux system call compatibility layer to the Windows NT kernel, and a Windows-native port of the popular “bash” shell. And furthermore, in partnership with Canonical, creators of the popular Ubuntu Linux distribution, the WSL supports Ubuntu’s “apt” package repository, giving Windows users access to a large swath of open source software built for Ubuntu, but running directly on Windows.
In short, what this means, is that Windows users will now have a Microsoft-supported Unix-like shell environment, and the ability to run Linux-based software directly on Windows, without a virtual machine. This would have been unthinkable to most even a decade ago.
Why porting UNIX software to Windows is hard
Software that is compiled from languages like C and C++, often favored by researchers, is generally built in such a way that the compiled binaries support a specific operating system. Each OS has a particular binary format–the way the program is organized on disk and copied into memory at runtime. So any compiled software built for that OS has to be arranged in the binary format for that OS in order for the OS to know how to interpret and execute it. It is not typical for one OS to be able to understand binaries for another OS. For example, software built for Linux uses the ELF binary format; normally if one tried to run a program built for Linux on Windows, which only understands the PE format, it will not be recognized as a valid executable.
An even deeper complication to writing portable software is the system calls– software run by users interacts with the operating system to perform low-level operations such as writing to disk, or making network connections, through special functions provided by the operating system called “system calls”. Modern UNIX-like operating systems follow, to an extent the POSIX standard for system calls, allowing them to be generally more interoperable. Windows, on the other hand, has its own system call defitions that are not necessarily in one-to-one correspondence with POSIX system calls. As such, a program built for Linux has no idea how to communicate with a Windows operating system.
This can be a problem even on higher-level interpreted languages like Python. Although code writing in Python abstracts away most operating system differences, Python code can still access OS-specific features such as system calls, and this is sometimes necessary to access more advanced OS features needed by some scientific software. So Python code that uses Linux- specific features, for example, can only run on a version of the Python interpreter built for Linux.
A third difficulty has to do with minor differences in user interface
standards. For example, a common issue in Windows support is its different
standard for representing file paths. While Windows paths contain a “drive
letter” and uses the backslash (“
\”) to separate between folders (e.g.
C:\Windows\cmd.exe), UNIX-like systems have no concept of a “drive letter”,
and use forward-slashes (“
/bin/bash). Issues like this can cause
many small, but pervasive bugs when porting software between operating systems.
How WSL gets around it
The Windows Subsystem for Linux does two main things:
It enables with Windows NT kernel to understand the ELF binary format, and translate it, as closely as possible, to the binary format used by Windows.
It implements a sizeable subset of the POSIX system call standard on top of Windows. Although Windows’ own system calls do not map directly the POSIX, because Microsoft has access to how its underlying operating system is implemented, they are able to implement the POSIX interface on top of the lower-level details of their NT kernel.
WSL also provides its own bash shell–a command-line interface favored by many users of Linux. This provides a UNIX-like command-line interface within Windows, also has an underlying system for transparently translating things like file paths between the Windows and UNIX formats.
The ultimate goal is to be able to take a program compiled and built on a Linux system, copy it over to Windows, and allow it to run without any modifications, with all the system-level translations completely transparent to the user. Targeting Linux software specifically makes this possible, because the system interfaces it will use are well-specified predictable in most cases. This is as opposed to running a virtual machine, in which an entire separate operating system is run in order to run software on that OS, and which needs to be able to run any arbitrary OS.
This is direct support for Linux software in Windows itself–there is no virtualization.
This is also an improvement over previous efforts at supporting Linux software on Windows, such as Cygwin. Because Cygwin is third-party software it cannot modify the Windows NT kernel itself. It does not support ELF binaries–to run software with Cygwin it has to be recompiled to the native PE binaries understood by Windows. It also does its best to provide emulation of POSIX system calls, but it has to do this by building them on to of the NT system calls which, as noted above, is not a one-to-one mapping. WSL, on the other hand, provides support directly from the operating system for POSIX and other Linux system calls.
What it means for OpenDreamKit
Because WSL allows binaries built for Linux to run directly on Windows, it makes much of the enormous repository of software built for Ubuntu (and potentially other Linux distributions) immediately available to run on Windows. No recompilation has to be performed or anything (at least, that is the goal–as we’ll see below it is still not fully realized).
For example, Ubuntu’s software repository already includes builds of many of the packages that are central to OpenDreamKit, such as GAP, PARI/GP, and some smaller packages including many of the dependencies of Sage. Sage itself has an unofficial Ubuntu package–this has been found so far to nominally “work” on WSL, but there have been found to be many bugs. That said, a great deal of other mathematical software–especially that which is less dependent on OS-specific features, should already work out of the box.
An additional potential advantage for WSL (indeed, one of the project’s goals as detailed in this article at Ars Technica) is to make the development tools and command-line interfaces favored by UNIX-oriented developers available on Windows. This makes it possible, in principle, to develop software like Sage the same way on both Windows and Linux.
In some sense this could be an end-run around OpenDreamKit’s goal of better supporting Windows–Microsoft has already done the lion’s share of the work for us. But there is more to be done, and it may not be an end-all be-all solution.
As mentioned in the previous section, while some OpenDreamKit software has been found to work in WSL, it is not without issues. Many bugs were found in running Sage on WSL (and even more when trying to compile it). This is not unexpected–the current release is marked “beta” by Microsoft, and they fully acknowledge that it is buggy and incomplete.
Second, Microsoft has made it clear in several statements, such as in this blog post that the WSL and “Bash for Windows” are to be considered tools for developer convenience only. It is not intended for use in a server infrastructure nor, presumably, as a means of distributing/installing software for end-users (i.e. who are agnostic about how the software is implemented). Although one could take the cynical view that this just Microsoft’s way of protecting its own server products, there are also some practical reasons for this:
As a developer tool, the WSL + Bash for Windows are not easy for casual users to install. First, it is only available on Windows 10 with the recent (as of writing) “Anniversary Update”. Not all users are on Windows 10 yet. It also requires having an account on Microsoft’s developer network, and for their Windows to be configured to “developer mode” in order to receive development-related updates, plus a few extra steps. This can also involve some sizeable downloads. This is not especially onerous for a developer, but is not a serious of steps that can or should be asked of the “casual” or first-time user just to install some software.
Despite having support directly in the kernel, the WSL is something of a walled garden. It is not possible to run native Windows applications from within the Windows bash prompt. Nor is it possible (in any transparent sense) to interact with Linux applications from native Windows applications. This is probably required, on some level, to maintain a clean abstraction.
Finally, it is not currently supported to run GUI applications on top of WSL, in part because that requires a lot more than just system call compatibility. While not supported officially by Microsoft, some hobbyists have made progress on it though, by integrating with existing X server implementations for Windows. For many mathematical softwares this is a non-issue–they are text based: numbers in; numbers out. Additionally, graphical interfaces for interactive research environments are increasingly moving to the web (see for example SageMathCloud). In such cases the GUI elements have been moved out to the web browser and the backend typically runs “headlessly”–it has no reliance on the system’s desktop interface.
The Windows Subsystem for Linux represents a major step in the right direction for Microsoft. It shows that they are listening to the needs of the broader software developer community (not just those who work exclusively on Windows) and that they have some interest in cooperating with the open source software community (this has also been demonstrated in several other ways in recent years).
For the purposes of OpenDreamKit, this work will make development of open mathematical software more accessible to a wider community. Although this may not improve accessiblity for casual end-users, many users of open research software tend to become de facto developers as well, as the more they use the software the more interested they become in modifying it for their own purposes. Making it possible for Windows users to do development on otherwise UNIX-oriented software, without leaving their personal desktop environments, is appealing. Being able to compile one’s own software is also important for some highly optimized numerical software, which tunes itself at compile time to the computer it is being built on, sometimes with dramatic results.
Although this does not yet provide a fully reliable immediate solution for porting OpenDreamKit software to Windows, we will continue to keep an eye on WSL as it evolves.