Re: [Axiom-developer] Axiom next release

[Camm Maguire]

Greetings!  (I'm away until 6/1, but am managing to read this remotely
today.  Unfortunately, communication with me might be unreliable until
6/1.) 



Mike Dewar <address@hidden> writes:

> Tim is right that compiling Axiom can take a long time but on a more
> modern machine and using CCL (which is *much* faster than gcc because it
> compiles to byte code and doesn't launch a C compilation for every file)
> it was taking us maybe 1 or 2 hours to rebuild all the algebra.  This
> didn't include the bootstrap process which you guys are presumably doing
> or any extra LaTeXing that the new pamphlets require of course.
> 
> I'd strongly recommend using CCL for your development system if you can.
> Not only is it the platform for the last NAG build but its smaller,
> leaner and as I've said compiles code much faster.  Although in theory
> akcl/gcl code runs faster in practice we found that you rarely got the
> benefit because the image was so much bigger and as a result the Axiom
> interpsys process swapped a lot more.  WIth CCL you can compile
> performance-critical parts of the libraries into the kernel (which
> roughly doubles the execution speed unless the functions are very small
> in which case the incresed overhead of calling compiled functions
> cancels this out), and as a result the first CCL-based version we
> shipped ran our test suite faster than the previous AKCL version.  Of
> course machines are bigger and faster today so this may not be such an
> issue any more.
> 

Just thought I'd note that, while I have no experience with Axiom and
its size issues, we've discovered with gcl/maxima and gcl/acl2 that
GCL in its default and traditional memory configuration gives up far
too much performance in an effort to save memory when running on
modern computer systems.  There is a brief note about this in the
recent 2.5.1 release notes.  Some users have found that by
preallocating a modest amount of memory in GCL, one can drive the
performance of certain maxima calculations to roughly three times the
speed of CMUCL, whereas by default, GCL is slower than CMUCL on the
same calculation due to its attempt to garbage collect too frequently
and save memory.  We plan on resetting the defaults in a later
release.

In any case, the summary picture appears to be that in today's world,
the performance/memory size vector points in the opposite way to that
traditionally expected.  

As I've stated before, I'm personally interested in seeing GCL be
useful for Axiom, and I hope that with its native object code
compilation, native object code linking into a running lisp core which
on most platforms can be saved to a system image for later use, 64 bit
support, 11 supported GNU/Linux architectures in addition to Sparc
Solaris and Windows, and fast developing ANSI Common Lisp support,
that we're already somewhat along down this path.

Take care,


> The biggest drawback with CCL was debugging the byte-code interpreter
> where neither a source-level tool looking at the C or the lisp-level
> tools gave us much of a clue as to what was going on. 
> 
> Well done getting integer arithmetic to work!  It sounds like you're
> making good progress.
> 
> Cheers, Mike.
> 
> On Tue, Feb 18, 2003 at 06:26:26PM -0500, Tim Daly wrote:
> > David,
> > 
> > There can't be a cycle of dependencies at the level of the Makefile.
> > I compile every file each time it is added to the Makefile and again
> > after I've added a set. Each file is independent, except at the
> > pamphlet level. So there can't be cycles in the algebra compile.
> > 
> > Axiom is huge. You're only building a portion of it so far. 
> > Non-linux compiles (or linux compiles on a machine with not a lot
> > of free memory) are going to take hours. Linux caches a LOT of the
> > disk transfers where Windows does not. The SUN build used to 
> > reload interpsys from disk every time as well as all of the loaded
> > files. Each algebra file compile would take 2 minutes or more which
> > gets into big numbers when you compile thousands of files. I've seen
> > an Axiom build on a SUN box take 24 hours even though I'd already
> > cached all possible work (hence the religious use of the intermediate
> > directory). I never built a SUN version from scratch.
> > 
> > Tim
> > 
> > 
> > _______________________________________________
> > Axiom-developer mailing list
> > address@hidden
> > http://mail.nongnu.org/mailman/listinfo/axiom-developer
> > 
> > ________________________________________________________________________
> > This e-mail has been scanned for all viruses by Star Internet. The
> > service is powered by MessageLabs. For more information on a proactive
> > anti-virus service working around the clock, around the globe, visit:
> > http://www.star.net.uk
> > ________________________________________________________________________
> 
> ________________________________________________________________________
> This e-mail has been scanned for all viruses by Star Internet. The
> service is powered by MessageLabs. For more information on a proactive
> anti-virus service working around the clock, around the globe, visit:
> http://www.star.net.uk
> ________________________________________________________________________
> 
> 
> _______________________________________________
> Axiom-developer mailing list
> address@hidden
> http://mail.nongnu.org/mailman/listinfo/axiom-developer
> 
> 
> 

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