2003 Kernel Summit: requirements panel

• Wim Coekaerts, Oracle kernel team.
• Walter Feldman, Unilever (which, he says, is moving large sections of its global operations to Linux).
• Bdale Garbee, recently named HP's Linux Chief Technology Officer.
• Robert Lefkowitz from "Big Unnamed Financial Institution" (the company did not want its name used at this event).
• Paul McKenney, IBM
• Tim Witham, Lab director, OSDL.

Each of the panelists was given an opportunity to present a wish list of desired kernel features. There was, as it turns out, quite a bit of overlap between their needs. Themes that came up include:

• Supporting large numbers of devices. At a minimum, 4,000 real devices need to be supportable on 32-bit systems. Linus pointed out that the x86 architecture is the real limiting factor; after a while, there is simply not enough low memory to keep track of that many devices. He is also not to thrilled by the recent 4G/4G patch (covered in the July 17 LWN Kernel Page), calling it "tasteless." This patch does make it possible to handle more devices (and memory), however, and Linus may eventually be prepared to merge it as an option. The real solution, though, is to "get off the drugs" and go to 64-bit systems if you need that many devices.

• There was also a request for support of truly large memory (up to 32GB) on 32-bit systems. That support is possible now - with the 4G/4G patch.

• Run-time storage reconfiguration without downtime. Moving disks and partitions around is a common operation, and they don't want to have to take the systems down to do it.

• Proper multipath I/O support. This item also relates to the "no downtime" requirement; the system needs to respond to I/O subsystem failures without going down.

• Distributed filesystems. Commercial offerings exist, but there is a real need for free, distributed filesystems. IBM has one in the process of being released, but IBM's lawyers are "busy with other tasks" at the moment. It was pointed out that Oracle has released its cluster filesystem under the GPL. Also available is OpenAFS, though its support in 2.6 is not what all would like. It was suggested that better support in the generic filesystem levels for clustered filesystems would be helpful.

• A cluster-aware logical volume manager is also on the list.

• A "more flexible" virtual memory subsystem. In particular, there is a desire for better control over memory usage by processes and process groups.

• The Big Unnamed Financial Institution representative spoke of a strong need for virtual machine support. Many of BUFI's applications require strict separation from each other; heavy use of virtual machines also makes testing of upgrades easier. Robert predicted that, in the near future, a substantial portion of deployed desktop Linux systems will actually be running under a system like VMware. User-mode Linux also works well for much of what BUFI is trying to do.

• Persistent device naming was mentioned by Bdale. When you have thousands of devices to manage, it really doesn't work to have them changing names frequently.

• Transparent superpages (support of large physical pages without the need for application changes). Large page support is available now, but it requires explicit calls from the application.

• Better support for performance counters. All of the existing performance counter implementations are either specific to a particular architecture, or to a specific application.

• Crash dump support. Vendors really do need this capability to diagnose problems encountered by their customers.

• Interrupt balancing. The hooks need to be put into place to allow user-space to query performance and set interrupt balancing policy.

• A process checkpoint and restart mechanism which eventually could grow to support migration of processes across systems.

• Asynchronous I/O for network operations.

• Standardized error logging which makes more sense to machines and humans alike. A more generalized, standard event reporting mechanism was also requested.

Linus reacted all of the above by asking about the needs of people with small systems - a group which makes up the majority of Linux users. Linux, he said, will not take over the very large systems arena at the expense of smaller systems. It was pointed out that many of the wishlist items are useful for smaller systems as well; an error and event notification mechanism can be used to inform the user of a low battery or full disk, for example.

The purpose of the session was to get the wishes out on the table, so few conclusions were reached. Much of this list will be revisited, in more detail, over the course of the summit.