[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Further speed comparisons of AKCL and Lucid


Please tell your independent-AI-company friends I do not claim that
KCL is faster than Lucid in general.  For each serious Lisp there are
probably benchmarks on which that Lisp is better than all the rest.  I
do claim that those who hear that KCL or ACKL is really slow ought to
take a careful look and do some benchmarking themselves.  On the tests
that I have run, AKCL and KCL are very competitive with the
alternatives I have looked at.  I have no financial interest in any
company that sells a Lisp system or Lisp hardware.

You are certainly right about how things can be tuned.  Take for
example the case of Moore's and my prover, which we recently converted
to Common Lisp.  When we completed the conversion, Lucid ran about 10%
or so faster than ACKL.  However, Bill Schelter made quite a few
improvements to AKCL and added some declarations to the prover, and
now ACKL does much better, as the following figures show:

    A comparison of nqthm times.  These times reflect all the standard
    nqthm provealls, including fm8501, but not including Goedel's theorem.

	       Time in  cpu seconds;  Normalized to akcl;   Time in cpu days

    akcl-280      43607.3                1.00                   .50
    akcl-75       74974.9                1.72                   .88
    kcl-280       80758.1                1.85                   .93
    symbolics1   101705.6                2.33                  1.18
    lucid-280    112875.1                2.59                  1.31
    symbolics2   151449.9                3.47                  1.75

    akcl-280, kcl-280, and lucid-280 were running on a 16mb Sun-3/280
    with a super eagle.  symbolics1, a 3600, had 4mb, IFU, Eagle.
    symbolics2, a 3640, had 4mb, no IFU, SCSI disk, output on Ethernet
    to a Sun.  Both Symbolics were running Release 7.1.  akcl-75 was run
    on froggy.ics.utexas.edu, an 8mb Sun-3/75, output on Ethernet
    to a Sun-3/180.

You will recall that I sent you the nqthm sources on the tape with
KCL, AKCL, and the version of the Gabriel benchmarks that we were
using, so you can run the nqthm tests, too.  I have no doubt that you
could change Lucid's Lisp to improve very significantly its
performance on these tests.  I would not be incredulous to hear that
an improved Lucid had managed to beat the AKCL times.  But note that
it would be hard for me or anyone else down here in Austin to make the
changes because I do not have the sources for Lucid, whereas we all
can get them for KCL and AKCL.  Thus the "tuning" issue is further
clouded by the "sources" issue.

As for your remark "I would expect a higher degree of effort from
someone making a public comparative analysis," please note that those
are cpu days in the table above.  I have been making comparisons of
Lisps since about 1975, I think, primarily to help me decide which
Lisp and which hardware to get.  I have a lot of interest in making
the right choice because it so affects my research.

As for your "junior-level" person's inability to produce via a
Sun-3/160 numbers comparable to mine and Schelter's for the Gabriel
times on rascal.ics.utexas.edu, our Sun-3/280, I offer you or this
person an account on rascal.ics.utexas.edu so you can check the
numbers we got, repeatedly.  Would you please send a copy of the Lucid
image that was being used by this person?  Perhaps that image was a
newer, better, or smaller version of Lucid than the one I was using.