On Sat, Nov 16, 2002 at 10:38:48PM +0100, Sapan Bhatia wrote: > Can anyone give me an idea of how significant the benifits are of > implementing network applications in the kernel as opposed to doing > so in userspace? > Judging by initial reviews of the in-kernel Tux webserver, which > supposedly outperformed Apache by upto 3 times fetching a mixture of > static and dynamic content, it really _does_ make a difference to do > implement things in the kernel. If this is true, then what kind of > overheads really make the differnce? I believe it was Ingo Molnar who wrote the x10 web server; entirely userspace but performance numbers equalling that of Tux. x10 would probably more popular except its license is a little funny.. That, and static web pages are just not interesting. The real difficulty is serving up dynamic web pages, and apache's flexibility is a real win here. Implementing services in the kernel is really only a performance win if you can avoid significant protection-boundary crossings. As you mentioned, the sendfile interface is pretty nice, and avoids most of the protection-boundary work involved in serving webpages. A kernel-based nfs daemon makes sense because it has more direct access to pages dirtied on the local machine. A kernel-based ipsec implementation makes sense because copying data from kernel buffers out to a process for decrypting and back across the kernel again for eventual delivery to another process is a lot of copies. But webservers have the keen sendfile interface to remove the largest amount of hassle. -- http://www.wirex.com/
Attachment:
pgp00213.pgp
Description: PGP signature
