Cygwin: A Free Win32 Porting Layer for UNIX Applications

6. Performance

Early on in the development process, correctness was almost the entire emphasis. As Cygwin became more complete, performance became a much important issue. We knew that the tools ran much more slowly under Win32 than under Linux on the same machine, but it was not clear at all whether to attribute this to differences in the operating systems or to inefficiencies in Cygwin.

The lack of a working profiler has made analyzing Cygwin's performance particularly difficult. Although the latest version of the library includes "real" itimer support, we have not yet found a way to implement virtual itimers. This is the most reliable way of obtaining profiling data since concurrently running processes aren't likely to skew the results. We will soon have a combination of the gcc compiler and the GNU profile analysis tool gprof working with "real" itimer support which will help a great deal in optimizing Cygwin.

Even without a profiler, we knew of several areas inside Cygwin that definitely needed a fresh approach. While we rewrote those sections of code, we used the speed of configuring the tools under Win32 as the primary performance measurement. This choice made sense because we knew process creation speed was especially poor, something that the GNU configure process stresses.

These performance adjustments made it possible to completely configure the development tools under NT with Cygwin in only ten minutes and complete the build in just under an hour on a dual Pentium Pro 200 system with 128 MB of RAM. This is reasonably competitive with the time taken to complete this task under a typical flavor of the UNIX operating system running on an identical machine.