Real-Time and Performance Improvements in the 2.6 Linux Kernel

The results in this section were compiled using a medium-strength Pentium-class system with a single 1.7GHz AMD Athlon processor and 512MB of system memory. The system was running the GNOME desktop environment and the system processes associated with the Fedora Core 3 Linux distribution, with up-to-date patches as of Feb 10, 2004. The system kernels tested were a vanilla 2.6.10 Linux kernel, the 2.6.10-1.760_FC3 kernel available as a Fedora Core 3 update, a vanilla 2.6.11-rc3 kernel and a 2.6.11-rc3 kernel with Ingo Molnar's current real-time and preemption patch. All of these kernels were compiled against the same kernel configuration file, modulo new configuration options introduced in the newer kernel sources.

In multiprocessing operating systems such as Linux, the system never is dormant. System processes such as the scheduler always are running. If you are using a graphical user interface (GUI), interfaces such as KDE, GNOME or standard X Window system window managers always are waiting for input events and so on. In order to examine true preemptibility and real-time performance, additional load was imposed on the system by starting various processes while each set of benchmark results were being collected. As mentioned previously, the system was running GNOME with four xterms open—one to run the realfeel benchmark, another to run a script that constantly ran recursive find and ls processes on the system's root partition and two in which 2.6.x Linux kernels, with separate source directories, were being compiled from a clean state.

Figure 1 shows a plot of the results of the Realfeel benchmark run on a stock Fedora Core system for a period of one minute. The system was running kernel version 2.6.10-1.760_FC3, which is a 2.6.10 kernel with various patches and enhancements applied by Red Hat. Each dot in the figure represents the jitter between an interrupt request and its handling. The X axis is the sample time in 1/60 of a second. Negative jitter numbers are displayed when the system responded to the interrupt faster than the projected standard time. As you can see from the figure, a fair number of these interrupt requests were handled exactly as expected, resulting in a visibly dark line along the 0 value of the Y axis.

Figure 1. Jitter Results on a Stock Fedora Core Kernel

Figure 2 shows a plot of the results of the Realfeel benchmark run on the same system with a vanilla 2.6.11rc3 kernel, which is release candidate 3 of the upcoming 2.6.11 kernel. These results also were collected over a period of one minute. As you can see from these results, the 2.6.11-rc3 kernel provides improved results from the FC3 kernel, with many more instances where the jitter between an interrupt request and its handling was zero.

Figure 2. Jitter Results on a Vanilla 2.6.11-rc3 Kernel

Figure 3 shows a plot of the results of the Realfeel benchmark run on the same system with a 2.6.11rc3 kernel to which Info Molnar's real-time/preemption patches have been applied. These results also were collected over a period of one minute, with the same load generators as before. As you can see from these results, the real-time/preemption patch provides impressively better jitter results, with relatively few departures from handling interrupts within the expected period of time. On the target system, these improvements translate into a much more responsive system, on which expectations about program execution are much more predictable than they are when running the vanilla FC3 or stock 2.6.11-rc3 kernels.

Figure 3. Jitter Results on a 2.6.11-rc3 Kernel with Real-Time/Preemption Patches

The improved scheduling, SMP and scalability improvements in the 2.6 Linux kernel provide higher performance Linux systems than ever before, enabling them to make better use of system resources and more predictably execute kernel and user tasks as requested by the system. Further improvements are available but currently are available only by patching your system manually or by obtaining a Linux distribution from a vendor such as TimeSys, which already incorporates and tests these high-performance patches.

The very existence of GNU/Linux as a free, open-source kernel and robust execution environment is something of a marvel. The contributions of individuals and, more recently, corporations to improving its performance will lead to an even brighter future. These and other improvements to Linux argue for and help guarantee the adoption of Linux as the preferred operating system for embedded, server and desktop applications.

Resources for this article: /article/8199.