|
From: <ebi...@ln...> - 2003-07-15 02:18:56
|
ja...@Mc... writes: > On Tue, 15 Jul 2003, Ken Yap wrote: > > > >Hmm, kinda makes me think of specifying a URL. > > >In fact, what if we could us HTTP to load a kernel ? > > > > > >Something like: > > > > > > filename = "http://192.168.0.254/lts/vmlinuz-2.4.21-ltsp-1"; > > > > > >Just thinking out loud :) > > > > You forget that HTTP uses TCP. Etherboot only supports UDP which is > > significantly simpler in implementation. But something I have thought > > about is a TFTP to HTTP proxy at the server side. One would have to pay > > some attention to security though. > > Ah yes, that little ole issue. > > So, is TCP really that much bigger/harder than UDP ? UDP is nothing but a port number so IP/TCP is certainly harder than IP/UDP. However the relevant case is how does: IP/UDP/TFTP or IP/UDP/NFS compare to IP/TCP/HTTP The big difference between IP/UDP/TFTP and IP/TCP for data transmission is that a lot more data can be in flight simultaneously. But I think a stupid TCP implementation would compare favorably with TFTP. However I have the strange feeling a IP/TCP/HTTP stack would be noticeably more complicated. On the wish list of IP features to implement there are also. - Fragmented packets. (Result in better TFTP throughput) - ICMP. (Closed port errors would be very useful to get) A very interesting question is what will happen when we finally hit the transition to IPv6. As DHCP is not necessary to get an IP address. We probably want to retain something like DHCP where we multicast for a server and then we get a unicast reply full of configuration information. The current design scales and load balances well. So nothing may change. But it may be the point where everyones newer and shinier implementations can be used. Eric |
