Torvalds Releases Linux 2.6.16 Kernel

Published: March 28, 2006

by Timothy Prickett Morgan

The wait is over, and the new Linux kernel, release 2.6.16, was put out by Linux creator and steward, Linus Torvalds, last week at kernel.org. While the support for the "Cell" Power-derived processor created by IBM, Sony, and Toshiba is interesting, what is more applicable to enterprise Linux customers is the improved support for NUMA clustering and support for new file systems.

Torvalds was brief in his announcement of the final release of the Linux kernel that will be at the heart of the commercial and private Linuxes that get packaged up for the next year or more. "Not a lot of changes since -rc6, but there's various random one-liners here and there (a number of Coverity bugs found, for example), and there are small MIPS and PowerPC updates," he said in his release note. "It looks like both Fedora and SuSE end up using a kernel that is pretty close to this 2.6.16 release, so let's all hope it's good. Give it a good testing, please."

You can view the changelog for the Linux 2.6.16 kernel here, and you can, of course, get the kernel and play around with it from kernel.org.

This is not, obviously, what the vast majority of enterprise Linux customers will do. They like packaged operating systems from Red Hat and Novell and a few other players. Now, the trick is to get the the new kernel rolled into the upcoming releases of SUSE Linux Enterprise Server 10, which is due in May, and Red Hat Enterprise Linux 5, which is due by year's end, as well as in the other releases from Turbolinux, Mandriva, the various Debian variants, and a few other regional players.

Sun Microsystems is obviously happy with Linux 2.6.16, since its patches to the kernel that allow it to support Linux on the new "Niagara" Sparc T1 processor made it into the final code. That means Sparc T1 support has a chance of becoming mainstream relatively soon as opposed to being a Linux variant that companies have to create on their own by compiling their own kernel. No potential Sparc T1 customer is going to compile their own Linux kernel, much less support such a compilation in a production environment. Similarly, as Sony and IBM try to build up the ecosystem for the Cell chip, support for Cell right out of the box is important if these companies want to drive the deployment of Cell in game machines (Sony) and in workstations and supercomputing visualization systems (IBM). As it is, the sophisticated co-processor design of the Cell chip, which has a Power5-derived core and eight special vector processors that are used to render graphics, is so different from a regular Power chip that it will require a significant amount of programming to make it useful. Having Linux support through the 2.6.16 kernel and, presumably, in future products from Red Hat and SUSE, means the Cell architecture can become something companies and academia can play with and then figure out how to deploy in real products.

While all of this is good, two changes that enterprises will be most interested in with the Linux 2.6.16 kernel are updates to the NUMA clustering algorithms and support for Oracle Cluster File System 2, which as the name suggests is a second generation of the cluster file system contributed to the open source community by database maker and application provider Oracle.

NUMA clustering is the way most high-end servers today are created, although most of the big iron server vendors still call it symmetric multiprocessing because early NUMA implementations had latency issues that made them relatively poor performers. NUMA stands for non-uniform memory access, and with a NUMA setup, you take a cell board from a server and carve up a piece of main memory to be local to that cell board and then another piece that can be accessed by any other cell board in the system. Since big servers based on Xeon, Power, Sparc, and now Opteron processors can have a dozen or more cell boards, the efficiciency and speed of communication between memory on one cell board and processors on a remote one is an important factor in the scalability of a server running a particular operating system and software stack. Unix machines have been NUMA-aware for many years, and Linux has been catching up quite nicely, as has the Windows platform from Microsoft. With the Linux 2.6.16 kernel, some changes have been made in the way this shared memory is allocated. If a processor is looking for data that is stored in the memjory on adjacent cell boards, the Linux 2.6.16 kernel can grab that information and move it over to local memory (which runs a lot faster) and then do the operation required on it without having to restart that operation. This is, in a way, a means to make NUMA clusters look more like a shared memory SMP cluster. Exactly who contributed this code to Linux is unclear, but this is one of the ways that Linux is going to be more scalable in the coming months.

Linux 2.6.16 also includes support for a similar protocol called TIPC, which is a message passing protocol that cell phone and switch maker Ericsson created for its switches and has contributed to the open source community to bolster the Carrier-Grade Linux project. TIPC can move small messages of under 1 KB between processors in a machine from 25 to 35 percent faster than TCP/IP and inside processes running on a single processor at 75 percent faster than TCP/IP. The connection mechanism in TIPC allows connections between processes to be initiated with as little as two messages, compared to a minimum of nine messages for TCP/IP links.

OCFS2 is a cluster file system that allows server clusters to share data stored on common storage area networks (SAN) or network-attached storage (NAS) arrays. It has been bundled in Novell's SUSE Linux Enterprise Server 9 since Service Pack 2 last summer. While OCFS2 is not bundled with Red Hat Enterprise Linux 4, since Red Hat has its own Global File System, it does run on this implementation of Linux. OCFS2 will run on any Linux 2.6 distro, but you have to go get it and do the work yourself. This is, again, not something most enterprises are interesting in doing. (You can find out more about OCFS2 at this link.) The Linux 2.6.16 kernel also has support for the Hierarchical File System for the Mac OS X operating system, which is called HFSX. What this will presumably allow is users on Macs to have a file system that Linux can also read.