A successful lisp machine?
Harvey J. Stein
Wed, 30 Apr 1997 02:40:33 +0300
Martin Cracauer writes:
> Harvey Stein wrote:
> > I thing the best bet is to:
> > 1. Start with Linux (or one of the free BSDs for that matter) + a
> > lisp or scheme.
> > 2. Write a loadable kernel module and/or the kernel patches needed
> > so that the lisp side doesn't have to fight the operating system.
> Do you want to put the Lisp system in the kernel or a LKM?
> Or just the "support" stuff such as memory tuning interfaces to
> synchronice paging and GC and or replacement for the siggnal
> messaging mechanism?
Well, I don't know about *me* putting it in... ;-). But I think it's
a fairly subtle problem to decide exactly what hooks need to go into
the kernel proper, what should go in as a kernel module, and what
should stay in user space. Some of the issues, as I see them, are:
-I suppose one needs to put enough in to make Lisp perform well and
interact well with libC, libraries written in C in general, and
the kernel system calls themselves.
-If the changes aren't *too* big, one might be able to get them
folded into the official kernel release.
-If the changes *are* big, one might be able to get a minimal set
of hooks into the official kernel release, and put most of the
needed code into a loadable kernel module. The dosemu people
have been known to do something along these lines.
-If everything that's needed can be done in a loadable kernel
module, then one doesn't need to convince any kernel developers
that one's changes should get folded into the source tree - just
distribute the module separately.
-Since kernel memory is nonpagable, one typically wants to do as
much as possible in user space so that memory can be better
managed. I hear that system call overhead in Linux is fairly low,
so this shouldn't imply much of a performance penalty.
In any case, I'd expect that
a) it'd be difficult to get a full Lisp system into the official
kernel sources, and
b) it probably wouldn't be so desirable anyway.
So, the answer to your question is that I was mostly thinking of using
a loadable kernel module for the support stuff, and begging the kernel
developers to add whatever additional hooks are needed by said
Harvey J. Stein
Berger Financial Research