[Gcl-devel] Re: killing off subordinate processes

[Camm Maguire]

Greetings!

OK, the latest fork modifications involving stack allocation in the
children is now checked in.  Doubtlessly need some cleanup and
documentation yet.  I want to try to be able to pass unreadable
objects before this is done -- the user should feel that there is
really just one (faster) lisp running when jobs are forked,
IMHO. Feedback on the user interface is most appreciated.  But the
basics of the parallelization I had planned for 2.7.0 are basically in
place.  I'd like to move to the remaining compliance issues and
compiler developments soon, as I will hav to suspend GCL work for some
time in mid Jan. and would like the release to be out and tested a bit
beofre then if possible.   To this end, bug reports and feedback along
the lines of those so helpfully offered thus far are always much
appreciated. 

PS, the fork overhead I now get on my box, which I believe is somewhat
slower than elgin, is 1 millisecond.  This will change somewhat as the
image churns itself from that present on the disk at system start.

PPS. setting si::*child-stack-alloc* to 0 passes all results through
print/read and can be used for comparison purposes.   This variable
represents the fraction of available stack space to use for
conventional allocation in the child, and defaults to 0.8.

=============================================================================
>(defun foo nil (si::p-let ((a nil)) a))

FOO

>(compile 'foo)

;; Compiling ./gazonk5.lsp.
;; End of Pass 1.  
;; End of Pass 2.  
;; OPTIMIZE levels: Safety=0 (No runtime error checking), Space=0, Speed=3, 
(Debug quality ignored)
;; Finished compiling ./gazonk5.o.
Loading /fix/t1/camm/debian/gcl/tmp/tmp/foo1/gazonk5.o
start address -T 0x844430 Finished loading 
/fix/t1/camm/debian/gcl/tmp/tmp/foo1/gazonk5.o
#<compiled-function FOO>
NIL
NIL

>(time (dotimes (i 1000) (foo))
)

real time       :      1.030 secs
run-gbc time    :      0.020 secs
child run time  :      0.080 secs
gbc time        :      0.000 secs
NIL

>=============================================================================

Take care,

Robert Boyer <address@hidden> writes:

> > using unwind-protect.
> 
> Thanks!
> 
> > this has added back another 10% or so.
> 
> sigh
> 
> Perhaps this will drive you to make unwind-protect faster someday!
> 
> Bob
> 
> 
> 

-- 
Camm Maguire                                            address@hidden
==========================================================================
"The earth is but one country, and mankind its citizens."  --  Baha'u'llah