Camm Maguire wrote:

> Greetings, and thanks for your report!
>
> Here is the patch you need:

> [...]

Thanks for the speedy reply, Camm -- the new patch did the trick for us.

> Do you find twc useful?  Any disadvantages for the generic lisp user
> that you can see?

I haven't seen any real disadvantages in practice.  I've gotten a couple of segfaults doing an ACL2 include-book, but such things are to be expected in "early access" code...if I try the same include-book  again, everything works fine.  In general I have noticed a reduction in GC activity with twc, but have seen a relative increase in the number of SGC's of STRUCTURE pages.  I'm not sure what STRUCTURE pages contain -- perhaps this is to be expected.

We have been able to build ACL2 (v 2.9), certify our own processor model (~100 books), and execute our CPU model code successfully using 2.6.6 twc (plus the aforementioned patch), which is a nice validation.  Some numbers for comparison (smaller numbers are better):

                       Before twc    With twc
                       ----------    --------
Certify 104 books      158 minutes   135 minutes
"Steady State" RSS     605 MB        409 MB
Execute CPU model      4.6 seconds   3.8 seconds
(small program)
Execute CPU model      333 seconds   268 seconds
(large program)

I think you'll agree that these results are quite encouraging!
 
We don't yet know what features of the twc code are most responsible for the performance improvements, but previous profiling (before twc) indicated that the processor model spent significant time performing
(expt 2 x) while running the "small program" mentioned above; this has been optimized in the twc builds.

Finally, note that we would rather be using 2.7.0, but experienced a show-stopping problem with t1 (yet another PROCLAIM error) -- we'll try t2, and let you know if the problem persists.


Thanks as always for your great work on GCL!

David Hardin
_______________________________________________
Gcl-devel mailing list
[email protected]
http://lists.gnu.org/mailman/listinfo/gcl-devel

Reply via email to