x-kernel

[Frans Kaashoek]

jj,

did you download the x-kernel sources?  i did it seven years ago (in
the good old days when i was a grad student).  what i remember is that
it made my understanding of the x-kernel much less abstract.  

on the design level, i think the x-kernel and click have quite
different goals.  the x-kernel focus is end-host protocol-stack
design.  they did some dynamic stuff, but that was on the level of "if
the other other party is on the same ethernet, we don't have run the
packet through the IP layer".  the x-kernel main abstractions where
headers and their main primitives popping and pushing headers from a
protocol stack.

the notion of microprotocols is an old idea, which Abbot and Peterson
developed further in an OO protocol language system---check the Prolac
paper.  however, if i remember correctly, that stuff couldn't deal
with complex protocols (too slow)---Eddie should have more details at
the top of his head.

    Frans

In message <pxhvh9aqtq1.fsf at grub.lcs.mit.edu>, John Jannotti writes:

>
>Now I see why people keep asking about x-kernel.  In their first paper, we
>seem quite different.  In their second ("A Dynamic Network Architecture",
>in ~jj/dynamic.ps)... well, there are a lot more similarities.
>
>I think we should say:
>
>We believe protocols should be be highly decomposed.  They begin to hint at
>"micro-protocols", but don't show any real work there.
>
>It's hard to compare performance (they are on a Sun3 on 10Mb/s ethernet),
>but our design seems lighter weight.  They associate each packet with it's
>own (lightweight) process.
>
>We focus on routers, not on demultiplexing to userspace.
>
>We make queues explicit, they seem to imply any faction could suspend a
>packet.
>
>We focus on making flexible routers mainly by rearranging factions, they
>seem far more willing to recommend faction modification.
>
>I *believe* we our system is better suited to reconfiguration without
>recompilation.  They are hazy about what exactly is involved in a new
>setup, but it *seems* to be a kernel recompile.
>
>
>  jj