Introducing the 2.6 Kernel

During 2.5, VM finally came into its own. The VM subsystem is the component of the kernel responsible for managing the virtual address space of each process. This includes the memory management scheme, the page eviction strategy (what to swap out when memory is low) and the page-in strategy (when to swap things back in). The VM often has been a rough issue for Linux. Good VM performance on a specific workload often implies poor performance elsewhere. A fair, simple, well-tuned VM always seemed unobtainable—until now.

The new VM is the result of three major changes:

The net result is superior performance in the common case without the VM miserably falling to pieces in the corner cases. Let's briefly look at each of these three changes.

Any virtual memory system has both physical addresses (the address of actual pages on your physical RAM chips) and virtual addresses (the logical address presented to the application). Architectures with a memory management unit (MMU) allow convenient lookup of a physical address from a virtual address. This is desirable because programs are accessing virtual addresses constantly, and the hardware needs to convert this to a physical address. Moving in the reverse direction, however, is not so easy. In order to resolve from a physical to a virtual address, the kernel needs to scan each page table entry and look for the desired address, which is time consuming. A reverse-mapping VM provides a reverse map from virtual to physical addresses. Consequently, instead of:

for (each page table entry)
      if (this physical address matches)
            we found a corresponding virtual address

the rmap VM simply can look up the virtual address by following a pointer. This method is much faster, especially during intensive VM pressure. Figure 4 is a diagram of the reverse mapping.

Figure 4. Reverse mapping maps one physical page to one or more virtual pages.

Next, the VM hackers redesigned and improved many of the VM algorithms with simplification, great average-case performance and acceptable corner-case performance in mind. The resulting VM is simplified yet more robust.

Finally, integration between the VM and VFS was greatly improved. This is essential, as the two subsystems are intimately related. File and page write-back, read-ahead and buffer management was simplified. The pdflush pool of kernel threads replaced the bdflush kernel thread. The new threads are capable of providing much-improved disk saturation; one developer noted the code could keep sixty disk spindles concurrently saturated.

Thread support in Linux always has seemed like an afterthought. A threading model does not fit perfectly into the typical UNIX process model, and consequently, the Linux kernel did little to make threads feel welcome. The user-space pthread library (called LinuxThreads) that is part of glibc (the GNU C library) did not receive much help from the kernel. The result has been less than stellar thread performance. There was a lot of room for improvement, but only if the kernel and glibc hackers worked together.

Rejoice, because they did. The result is greatly improved kernel support for threads and a new user-space pthread library, called Native POSIX Threading Library (NPTL), which replaces LinuxThreads. NPTL, like LinuxThreads, is a 1:1 threading model. This means one kernel thread exists for every user-space thread. That developers achieved excellent performance without resorting to an M:N model (where the number of kernel threads may be dynamically less than the number of user-space threads) is quite impressive.

The combination of the kernel changes and NPTL results in improved performance and standards compliance. Some of the new changes include:

The results speak for themselves. On a given machine, with the 2.5 kernel and NPTL, the simultaneous creation and destruction of 100,000 threads takes less than two seconds. On the same machine, without the kernel changes and NPTL, the same test takes approximately 15 minutes.

Table 4 shows the results of a test of thread creation and exit performance between NPTL, NGPT (IBM's M:N pthread library, Next Generation POSIX Threads) and LinuxThreads. This test also creates 100,000 threads but in much smaller parallel increments. If you are not impressed yet, you are one tough sell.

Table 4. Results of the thread creation and exit test: this test measures the time for ten initial threads to each create and destroy one, five or ten parallel threads.