BackgroundThe project I’m working on is written in C++, with some Python tools mixed in. My original development environment was Visual Studio 2005 on Windows XP. This is already the first issue: Updating Visual Studio or Windows is not trivial, as both the OS upgrade as well as IDE updates require new licenses, and especially in companies new versions are not bought immediately.
The problems became apparent when I tried to multi-thread parts of the application. At the core, it’s doing a lot of number crunching, in small work blocks which can be processed independently. As I couldn’t use OpenMP due to dependency issues (a 3rd party library could not be linked when OpenMP was enabled), I was threading manually.
Unfortunately, the application had to allocate some memory in each thread, and as it turned out, the scaling on XP was catastrophic. While I did get a speedup from 1->3 cores, it became slower from 3->4 – clearly, I was hitting some issues with either the scheduler or the memory subsystem, as my code didn’t have any I/O in it.
A quick try on Vista showed that the same application ran more than twice as fast, but unfortunately, I couldn’t install my IDE on Vista as well, and developing on Windows XP and testing on Vista was out of question. Again, with a free IDE, the change would have been no problem (and the express editions don’t have x64 support*, nor OpenMP, more on this later. * See below for an update!).
On the other hand, getting a recent Linux with a recent compiler was not a problem. With Wubi, Linux can run side-by-side with Windows while giving you a full Linux based development environment. Running side-by-side is especially important when you’re in a corporate environment, as you usually cannot simply erase the disk and install Linux without making the IT angry.
LinuxI used Wubi, with the Kubuntu flavour, as I like the KDE environment a bit more than GNOME – especially as I use Qt for UI development now. Specifically, I used Kubuntu 9.04 x64, while I used a x86 Windows XP previously.
PortingI started by checking out the code from SVN, so no problems here. Even though the application was written in standard C++ and didn’t depended on Windows-specific functionality, it was built using Visual Studio project files and used a few WinAPI calls.
As a first step, I ported everything to CMake – something I could/should have done on Windows already. With CMake, I was able to quickly convert one project after the other, and immediately check for compile errors. This proved to be the best way, as I never came into the situation where I would get huge amounts of compile & linker errors at the same time; I had that when I moved an already CMake based project from Windows to Linux and tried to get it running in its entirety. During porting it’s best to port on subproject at a time, even if the project is originally using a portable build system.
As I mentioned, I used explicit threading on Windows, which I replaced by OpenMP on Linux. Now I could also throw out all configuration stuff for threading; among other things, I wouldn’t have to reduce the priority of my application on start-up – this was necessary on XP, as the machine would become unresponsive during processing otherwise.
Boost.Threads might have been a valuable alternative here, but OpenMP is well suited to the kind of loop-parallelism I had in my code, and even simplified it compared to the explicit splitting/execution I used previously.
For graphics, I was already using OpenGL. As I could easily get the nVidia binary drivers running, I had no trouble on this side. Overall, it took me half a day to port, including the time to set up the Linux installation.
Results and some thoughtsThe net result is interesting: The same application is running 5-10x faster now when using all four cores, so porting to Linux was really worth the hassle. I assume that with Visual Studio 2010, running on Windows 7, I would get similar performance, but the key point to take with you here is: Getting your stuff to work on Linux only costs you time, and not too much if you are a bit careful.
Using CMake on Windows (or another portable build system), writing more or less clean C++ and using portable libraries makes porting to Linux easy, and the switch itself is not too complicated. At the moment, the tooling on Linux is reasonably comfortable (more on this in the next post), and the “pain factor” to switch from Visual Studio to for instance KDevelop or Eclipse CDT is no longer there.
Actually, the switch is so simple that Microsoft should get concerned. For instance, I have been developing mainly on Windows since several years, and I occasionally tried Linux, but I never did a complete switch due to various smaller and bigger problems.
However, since 1-2 years, the Linux desktop, together with the tools, is good enough to provide some real benefit, especially if you cannot access the latest Microsoft products. Microsoft used to have the best developer tools by far, and quite stable APIs, which were in my opinions the corner stones of their success.
However, they’re changing APIs now rather quickly (WinForms? WPF? WinAPI?), they provide new platforms which require rewriting your applications (I’m still waiting for an application like AutoCAD which has a C# UI and a C++ backend), and the tool release cycle is simply too long – waiting 2 years to get a compiler bug fixed is just ridiculous.
On the other hand, developing on Linux means you have an extremely stable API (POSIX isn’t going to be replaced with PoseFX, for instance), the UI side is rather clear (GTK or Qt, you choose), and the tools are getting better as well (GCC and LLVM are getting better quickly, and installing a new GCC does not require buying a new license).
If Microsoft does not turn around the ship with Visual Studio 2010 and some clear statements on the APIs, I assume that more and more developers will find that Linux can be also a very nice environment. Again, more on this next time!
[Update] A few notes, as this article is getting a lot of attention and there are some misunderstandings. Regarding the library, I had access to the source, and I could have built it on Windows — it’s simply very time-consuming, as it depends on many libraries like zlib, which I would have to compile with the new settings on Windows as well. Getting these dependencies on Linux is obviously much easier.
OpenMP vs. manual threading:
This did not improve the performance compared to the manual threading, but it was a nice bonus as it cleaned up the code a bit. On Windows, I could not use OpenMP due to link issues. The other bonus came from the switch from x86 to x64.
Finally, the Linux memory allocator is much better in multithreaded environments, and this was the biggest performance improvement. In total, the application runs several times faster now, without changes on my side.
Summary: In this particular case, the switch to Linux took me just a few hours, while the benefit for me was rather big: Improved performance, and less hassle with dependencies. The drawbacks — getting used to Linux, different tools — are out weighted by the advantages for me; and that’s the main point of this blog post.
I was surprised how easy it was to switch from Windows to Linux completely on this project; as I expected a lot of problems (like for instance, not being able to get the stuff running at all!)
Things might have been different if the library and the dependencies would be easier to build on Windows, and I would have gotten access to a new Windows version/new compiler version, but this is simply how things turned out, and I don’t miss my Windows setup at work.
[Update] The x64 editions of Visual C++ Express require some setup to support x64, but it is possible to add x64 support manually using the compilers from the Windows SDK. Check out the official C++ Express feature list and a guide how to enable x64.