next up previous
Next: Ports Up: Performance improvements Previous: Block driver

Native kernel

In order to get the best performance with this port, some work will likely be needed in the host kernel. The need for system call interception without context switches has already been mentioned. In addition, access to the host memory context switching mechanism would probably speed up context switches greatly. The ability to construct and modify mm_struct objects from user-space and switch an address space between them would eliminate the potential address space scan from context switches.

Another area to look at is the double-caching of disk data. The host kernel and the user-mode kernel both implement buffer caches, which will contain a lot of the same data. This is obviously wasteful, and tuning the host to be the best possible platform will probably require that this be addressed somehow.

Jeff Dike 2000-08-25