Re: [Gcl-devel] hash redux

[Matt Kaufmann]

Hi, Jared and Camm --

I ran the experiment you suggested, Jared (thanks for the suggestion).
In books/centaur/gl/:

(ubt! 1)
(include-book "portcullis")
(rebuild "solutions.lisp" t)
(u)
(time$ (def-gl-thm 1f
  :hyp (and (unsigned-byte-p 3000 x)
            (unsigned-byte-p 3000 y))
  :concl (equal (+ x y) (+ y x))
  :g-bindings (gl::auto-bindings (:mix (:nat x 3000)
                                       (:nat y 3000)))))

That took 78 seconds (a very nice improvement!).  Then:

ACL2 !>:q

Exiting the ACL2 read-eval-print loop.  To re-enter, execute (LP).
ACL2>(hons-summary)

Normed Objects Summary

 - NIL-HT:                     4 count,           5,000 size ( 0.08% full)
 - CDR-HT:             9,071,974 count,      12,974,622 size (69.92% full)
 - CDR-HT-EQL:                 0 count,           1,000 size ( 0.00% full)
 - STR-HT:                     1 count,           1,000 size ( 0.10% full)
 - PERSIST-HT:                 0 count,             100 size ( 0.00% full)
 - FAL-HT:                     0 count,           1,000 size ( 0.00% full)

NIL

ACL2>(hl-hspace-str-ht *default-hs*)

#<hash-table 0000000004e06af0>

ACL2>

(I did some searching and did find another 'equal hash table besides
that str-ht, namely; *hcomp-book-ht*, but it's quite small and not
relevant here.)

So I'm again stumped, since the cdr-ht is, I think, an 'eq hash
table.

Camm, is there a way to identify the callers that are setting a hash
table with test 'equal?  The profile you sent seems to be at the level
of C, so I don't know what to trace.

-- Matt
   From: "Jared C. Davis" <address@hidden>
   Date: Mon, 11 Nov 2013 16:18:24 -0600
   Cc: Camm Maguire <address@hidden>,
           "address@hidden" <address@hidden>

   Hi,

   I believe Matt is correct that the only use of EQUAL hash tables in
   the (h) part of ACL2(h) is for string hashing.  In fact, for the most
   part, in a single-threaded context, I think there should typically be
   just a single string hash table.

   At the relevant part of your benchmark, you might run (hons-summary)
   to see the size and count of this table, in case that's helpful.  Or
   if you want to get your hands on the hash table to really take a deep
   look at it, you can try, e.g.,:

   ACL2 !>(hons "foo" "bar")
   ("foo" . "bar")
   ACL2 !>:q
   :q

   Exiting the ACL2 read-eval-print loop.  To re-enter, execute (LP).
   ? (hl-hspace-str-ht *default-hs*)
   #<HASH-TABLE :TEST EQUAL size 2/1000 #x30200EA5441D>

   Cheers,
   Jared

   On Mon, Nov 11, 2013 at 3:21 PM, Matt Kaufmann <address@hidden> wrote:
   > Hi, Camm --
   >
   > That's interesting, but I'm confused, and I'm definitely not an expert
   > on hash tables.  I looked at the files that implement the "(h)" part
   > of ACL2(h), which is almost certainly what is involving hash tables,
   > and it looks to me like maybe the only 'equal hash tables are for
   > strings.
   >
   > I'm forwarding this to Jared, since he is the most recent author of
   > that code (plus, you mention him as helping with potentially related
   > reader issues), in case he has time to shed light on this.
   >
   > -- Matt
   >    From: Camm Maguire <address@hidden>
   >    Date: Mon, 11 Nov 2013 12:49:41 -0500
   >
   >    Greetings!
   >
   >    Just a followup -- the remaining time appears to be in sethash for an
   >    'equal hash-table:
   >
   >    
=============================================================================
   >    index % time    self  children    called     name
   >                                 103979625             sethash [1]
   >    [1]     84.2    2.11   49.03       0+103979625 sethash [1]
   >                   22.58    6.16 167566772/167566772     fShash_equal [2]
   >                    0.00   20.28  119656/131885      alloc_relblock [6]
   >                    0.01    0.00  119656/205048      alloc_object [47]
   >                                 103979625             sethash [1]
   >    -----------------------------------------------
   >                   22.58    6.16 167566772/167566772     sethash [1]
   >    [2]     47.3   22.58    6.16 167566772         fShash_equal [2]
   >                    5.25    0.00 363849475/363849475     hash_eql [12]
   >                    0.91    0.00 1174935219/1174940911     eql1 [18]
   >                    0.00    0.00      12/2577623     Fand <cycle 2> [151]
   >    -----------------------------------------------
   >                    0.29    5.78       3/14          make_cons [9]
   >                    1.06   21.19      11/14          alloc_relblock [6]
   >    [3]     46.6    1.35   26.97      14         GBC [3]
   >                   26.93    0.00 25304834/25331171     sgc_mark_object1 
<cycle 1> [5]
   >    
=============================================================================
   >
   >    This is somewhat remarkable, as the 'eql gethash calls which greatly
   >    dominate in number are no longer on the radar.  Presumably the algorithm
   >    makes some complex cons, (definitely not your grandmother's '(1 2 3)
   >    list), uses an 'equal hashtable to make it equal-unique, and then uses
   >    that as a key in an 'eql hashtable for the real heavy work.
   >
   >    This just reminded me of the work we did earlier regarding the loading
   >    of complex conses in compiled files, which overloaded the #= reader
   >    until we memoized the routine calculating the hash-equal index.  This is
   >    barely necessary to the gcl compiler -- the point is to catch errors
   >    where the constant list to be compiled in changes during compilation.
   >    And as I indicated earlier, we flush the memoizing hash tables on each
   >    compile-file.  This, together with the implementation of the 'hybrid' #=
   >    algorithm suggested by Jared, made the loading of these conses very
   >    fast.
   >
   >    My question is if we've learned anything which might make the above
   >    results yet faster.  By default, the hash-equal index descends no more
   >    than three levels, car and cdr, into a cons to xor up the index.  It
   >    does not attempt to descend the entire structure memoizing as one goes
   >    like the compiler version.  There the depth limit is much greater (1000)
   >    due to its purpose and the absence of any table.  My intuition tells me
   >    that there is no way a memoized version of the generic hash-equal would
   >    pay off.  It seems we would have to flush on each call, or never.  It
   >    would only speed up index calculations of great depth, which is only
   >    useful in hash tables if your index is insufficiently random at the
   >    default depth of 3.  This does not appear the case, as #'equal itself is
   >    absent from the profiling report, implying that the hit rate to the
   >    index is good.
   >
   >    I suppose an 'equal hashtable could keep an 'eq hashtable internally for
   >    the life of the table.  That might be interesting.
   >
   >    In any case, I don't want to waste a lot of time reinventing some
   >    wheel.  If you or any of the other hashtable experts have some wisdom
   >    here, I'd be most appreciative.
   >
   >    Take care,
   >    --
   >    Camm Maguire                                        address@hidden
   >    
==========================================================================
   >    "The earth is but one country, and mankind its citizens."  --  
Baha'u'llah
   >
   >
   > _______________________________________________
   > Gcl-devel mailing list
   > address@hidden
   > https://lists.gnu.org/mailman/listinfo/gcl-devel



   -- 
   Jared C. Davis <address@hidden>
   11410 Windermere Meadows
   Austin, TX 78759
   http://www.cs.utexas.edu/users/jared/