CS262B Reading Summary

Extensible Kernels are Leading OS Research Astray

Peter Druschel, Vivek S. Pai, Willy Zwaenepoel

Feng Zhou
1/25/2004

Valid points of the paper are:

1. This interesting short paper argues that extensible kernels will never make into the mainstream OSes because they bring only marginal benefits and a whole lot of hard issues to a production system.  The issues include performance, evolution and safety.  A lot of performance gains reported by previous extensible OS research can actually be implemented in non-extensible OSes as well.  On the other hand, extensible OSes are hard to evolve because they widen the interface between the kernel and the rest of the system.  Solving the safety problem introduces new problems like type-safe languages and trusted compilers.
2. The paper divides extensions to a kernel into several categories, namely enhancements, inherently safe customizations and general extensions.  Only general extensions are not implementable in a traditional OS.  The authors argue that almost all useful changes to the OS belongs to the first two categories, i.e. a "right interface" can almost always be found for a certain kind of functionality.  Therefore, extensible kernels have their use in the labs, but not on production systems.  They should be used to experiment with the kernel and help finding the "right interface".

The IO-Lite example is discussed too succinctly.