One of the most successful mass-media campaigns of the last few years was Volkswagen's "Drivers Wanted" campaign, which built VW's image as a kind of junior BMW. Those of you who are interested in more admercial stuff probably picked up the wrong magazine. What I'm talking about here is the pressing need for tuned, server-class device drivers in Linux.
Until recently, there was always a good scapegoat for the limited performance of high-end Linux device drivers: the kernel. The technical limitations of the 2.2 kernel, when applied to the unforgiving real world of high-performance IT, meant that Linux rarely got deployed in situations where a Fibre Channel driver was needed to drive devices at 80% of their rated capacity. It seemed that drivers for this class of hardware were developed to the point of confirming that they seemed not to crash and then the developers moved on to other problems.
Now that the new 2.4 kernel is being rolled out, customers are really planning to deploy Linux systems in big-time applications that have been dominated by the likes of Sun Solaris. The scalability for Linux will now definitely be there for high-end scenarios. On the SMP front, Linux will see much greater scalability in terms of number of processors. The TCP/IP networking stack has been streamlined for the next wave of Gbit networking. And developments like ReiserFS are toughening up the ability of Linux to handle much larger data sets.
There is, however, one storm cloud still lingering on the horizon. While the Linux kernel has been revved up for greater throughput, old drivers that were once an annoyance will now become colossal bottlenecks. With the current crop of device drivers, getting data into and out of a Linux system will be slower and far more painful than it should be.
The need is clear, but the solution is not as simple as it should be. That's because today the best Linux-driver development is done in academia and on home systems. Now, let's face facts: How many of those people have a spare Fibre Channel or gigabit Ethernet card handy? So is it any surprise that the best work goes into ATA and video drivers?
Fortunately, there are a few hardy souls who concentrate on high-end device-driver development. Recently, I had a chance to ask one of the best, Matt Jacobs, what he thought needed to change to make Linux driver development happen faster. He believes that many problems stem from a lack of true coordination of work among developers in different parts of the community.
One very sensible suggestion he made was to have some kind of architectural roadmap for driver development as Linux moves forward.
Documenting this would empower a wider circle of talent than the word-of-mouth group that currently does so much of this important work. A peer-review process, the use of network revision-control tools like CVS, and better synchronization of releases for different hardware platforms would also help tremendously, he believes.
I would also suggest that better high-end drivers will only be developed when the vendor community starts to take matters into its own hands. Vendors need to try to identify people doing the good work and provide focused sponsorship of that work with loaner hardware and subsidies as needed. We also need to look at getting more targeted Open Source in-house efforts funded. If that happens, Linux drivers will improve as fast-or perhaps even faster-than the overall system. Only in that way, whether we drive VW s or not, will we be able to whistle past Microsoft and see Sun in our rear-view mirrors.
-Sam Bogoch can be contacted at sbogoch@bigstorage.com
| 
