Implementing protocols in user space versus kernel space - performance implication

Question

I am looking for some articles / papers on the impact of moving a protocol implementation say TCP/IP stack from kernel to user space. It will obviously have an impact but how much? Or there is any literature of how much the context switch costs. I realize that there may not be an accurate answers because it will depend on the application. Assuming therefore it is a network stack, it will be great if anyonce can probvide some inputs. I googled but cannot find anything good except this User space Vs Kernel space program performance difference but which does not shed enough light.

Answer 1

As I know you could see for file system comparisions - there is many different implementations, as kernel modules and as fuse modules. May be you can find interesting information in this types of components.

Also in field of network I can mention that some drivers utilize ability of network cards to calc hash's for ethernet frames - and there is some gane of it. If you work on user space - you can't use such trick. In result you can get compassions of performance in use of such trick as upper bound of performance penalty.