[Openmcl-devel] OpenMCL Versions

[Gary Byers]

On Wed, 16 Aug 2006, Andrew Shalit wrote:

> If MCL was targeting Windows desktop applications, Intel 32 would be
> more important.  But I'm not hearing that from anyone, and looking
> forward the 64 bit architecture clearly is the way to go.  I also like
> your point about multiple cores.  I don't know whether Gary has put
> thought into that.  Gary?
>

The lisp generally has no idea of how many CPUs/cores are available;
keeping CPUs busy is the OS's job.  The more cores are availabe,
the more threads can run concurrently; the things that have stressed
OpenMCL the most probably have more to do with having lots of threads
running (increasing the chance that the GC will stop a thread when
it's in some state that the GC misinterprets.)

David Rager's done a lot of work on parallelizing funtional programs
in OpenMCL (and that work exposed a lot of GC timing screws and other
issues), and his system uses the number of available CPU cores as a
parameter (tries, IIRC, to keep the number of runnable threads 
proportional to the number of cores.)  I -think- that once we got
past the worst of the "GC screws and other issues", he generally
obtained good scalability in cases where he expected to (but he
may remember differently ...)  He was developing (and I was debugging)
on very similar Quad G5 systems, and I'm fairly sure that I've seen
"top" claim that the dppccl or dppccl64 process was approaching
"400% CPU usage" in some cases.