[Monetdb-developers] hashjoin and strHash

Wouter Alink wouter.alink at gmail.com
Sat Dec 19 12:21:09 CET 2009


Lefteris, you are correct in that i meant 'second time the query was
run' when I wrote 'hot run'.

I see that at GDK level reuse cannot be estimated. Although with
current hardware which has an abundance of memory, and the fact that
strings take up much more storage than a single BUN (so a hash-entry
is usually relatively small compared to its data) GDK might weigh the
additional costs. GDK also decides which things to keep in memory or
throw it out, which in turn is also based on reuse.
The costs for performing the initial join are dominated by the strHash
function, and building the hashtable on the big BAT or the smaller BAT
makes (almost) no difference, except for the additional memory use. If
on such a big bat again a join is performed, it will be beneficial to
have the hashtable in place.

What I was hoping for were explanations of situations where it makes
no sense to build the hashtable on the bigger string BAT, but a good
counter-example I haven't seen. In general, i can see, it would not be
beneficial if the big BAT is not joined twice, but if it doesn't hurt
too much, couldn't it just be the default?

Eventually I would like to be using the SQL layer only. Here there
would be plenty of tables with string-columns, and some will be joined
against. Should a MAL optimizer detect that I am about to join two
string-BATs, and that one BAT is bigger than the other and has many
different values, and therefore should build a hashtable on the bigger
one? The MAL optimizer can only guess about my next query (although I
agree that it could do a better job at guessing), and calculating
heapsize/batsize seems to be an operation that is also difficult to do
on a MAL layer.

is really nobody in favour of changing the behavior of joining string
bats for large bats with many different values? well, than I give up.
Wouter


2009/12/18 Stefan Manegold <Stefan.Manegold at cwi.nl>:
> Hi Wouter,
>
> in the lines of Lefteris' reply:
> for a single join with no hash table present a priori, the number of hash
> function calls is euqal to the number of BUNs in both BATs; for each inner
> BUN the function need to be called to build the hash table, for each outer
> BUN it needs to be called to probe the hash table. Hence, the pure hashing
> costs are independent of which BAT is inner and which outer.
> Given that, the reason the choose the smaller as inner is indeed to increase
> spacial locallity (and thus performance) of the inherently random access
> while building and probing the hashtable.
>
> As Lefteris pointed out, the "operational optimization" in GDK is a pure
> peephole optimization dealing only with the very operation at hand. I.e., in
> general it cannot anticipate the benefits of future re-use of efforts, like
> investing in the (more expensive) building of a larger hash table to be able
> to re-use this in several later operations --- which IMHO is independent of
> the data type. Such descisions need to be made at higher levels, either in
> MAL optimizers or in the front-end that generates the MAL plan.
>
> Stefan
>
>
> On Fri, Dec 18, 2009 at 05:01:07PM +0100, Lefteris wrote:
>> Hi Wouter,
>>
>> funny think, I had the same exact problem and we were thinking about
>> this issue. The idea here is that this observation for strings might
>> not be always true, and it is a situation that cannot be always
>> determined on the kernel level. Correct me if I am wrong, but your
>> benefit on query comes because the hash in the large BAT is already
>> there, that's why the second time you get 0.01? You mention hot run so
>> I assume the BAT is already there with a hash index. While in the
>> original situation the hash is on the small BAT thus you don't benefit
>> from the hot run. But if a big BAT of strings is to be used again it
>> is unknown in the gdk level. So, I solved the problem by forcing the
>> hash index on the big BAT in a higher level (in Monet5) where it knows
>> something more about the application (in my case RDF store). Can you
>> do instead that? force the hash index in a higher level for you
>> application? If gdk see a hash index already there, then it will
>> choose that independent of the size.
>>
>> lefteris
>>
>> On Fri, Dec 18, 2009 at 4:22 PM, Wouter Alink <wouter.alink at gmail.com> wrote:
>> > Dear developers,
>> >
>> > I would like to propose a change in GDK and hear opinions. It is about
>> > the following issue:
>> >
>> > in the BATjoin code, if there is no possibility to do a fetch or merge
>> > join, a hashjoin is performed. A hashtable is created for the smallest
>> > BAT. The reasons (i could think of) for choosing the smallest BAT for
>> > the hashtable are that less space is required for the hashtable (which
>> > in turn causes less cache misses when doing a lookup) and also because
>> > the hashfunction used is assumed to be very inexpensive (it needs to
>> > be calculated for each item in the large bat each time a join is
>> > performed).
>> > I can see that the hashfunction can be very efficient for data types
>> > without indirection, but I feel that for data types like strings in
>> > some cases this is a little different. If a string BAT for example
>> > contains many different values (i.e. is not a bat which contains
>> > enumeration values) the hashfunction will not be inexpensive anymore
>> > (many cache misses), as each hashfunction call needs to hash a whole
>> > (arbitrary long) string at an arbitrary place in the heap.
>> >
>> > Is it perhaps possible to specify that, when a BAT of type 'str' has
>> > many different values a hashtable may be build on the large BAT
>> > instead of on the small BAT?
>> >
>> > Reason that I ask this: I was analysing costs of a query in which I
>> > had a few short strings (26 tuples, 1-column table: varchar) which I
>> > wanted to look up in a dictionary (9M tuples, 2-column table:
>> > int,varchar).  "SELECT a.id FROM longlist AS a JOIN smalllist as b ON
>> > a.strvalue=b.strvalue;"
>> > The result is a small list of integers (26 or less tuples). This
>> > operation currently takes roughly 1.5 seconds for a hot run, mostly
>> > due to 9M strHash operations. By applying the patch below the
>> > execution time for a hot run dropped down to .01 seconds. The
>> > performance gain is caused by only having to perform strHash on the
>> > items in the small bat once the hashtable for the large bat has been
>> > created.
>> >
>> > Any suggestions whether such a change is useful? Which benchmarks will
>> > be influenced?
>> >
>> > I guess this code change is probably not useful for large string BATs
>> > with only few different values, but perhaps a guess could be made how
>> > diverse the strings in a bat are (by taking a sample or perhaps simply
>> > by looking at the ratio batsize/heapsize), and based on that determine
>> > whether to build it on the large or small BAT?
>> >
>> > Greetings,
>> > Wouter
>> >
>> >
>> > Index: src/gdk/gdk_relop.mx
>> > ===================================================================
>> > RCS file: /cvsroot/monetdb/MonetDB/src/gdk/gdk_relop.mx,v
>> > retrieving revision 1.167.2.4
>> > diff -u -r1.167.2.4 gdk_relop.mx
>> > --- src/gdk/gdk_relop.mx        20 Nov 2009 13:04:06 -0000      1.167.2.4
>> > +++ src/gdk/gdk_relop.mx        18 Dec 2009 14:59:13 -0000
>> > @@ -1232,7 +1232,12 @@
>> >  @-
>> >  hash join: the bread&butter join of monet
>> >  @c
>> > -       /* Simple rule, always build hash on the smallest */
>> > +       /* Simple rule, always build hash on the smallest,
>> > +                except when it is a string-join, then we do the opposite */
>> > +       if (swap && rcount < lcount && l->ttype == TYPE_str) {
>> > +               ALGODEBUG THRprintf(GDKout, "#BATjoin:
>> > BATmirror(BAThashjoin(BATmirror(r), BATmirror(l)," BUNFMT "));\n",
>> > estimate);
>> > +               return BATmirror(BAThashjoin(BATmirror(r),
>> > BATmirror(l), estimate));
>> > +       }
>> >        if (swap && rcount > lcount) {
>> >                ALGODEBUG THRprintf(GDKout, "#BATjoin:
>> > BATmirror(BAThashjoin(BATmirror(r), BATmirror(l)," BUNFMT "));\n",
>> > estimate);
>> >
>> > ------------------------------------------------------------------------------
>> > This SF.Net email is sponsored by the Verizon Developer Community
>> > Take advantage of Verizon's best-in-class app development support
>> > A streamlined, 14 day to market process makes app distribution fast and easy
>> > Join now and get one step closer to millions of Verizon customers
>> > http://p.sf.net/sfu/verizon-dev2dev
>> > _______________________________________________
>> > Monetdb-developers mailing list
>> > Monetdb-developers at lists.sourceforge.net
>> > https://lists.sourceforge.net/lists/listinfo/monetdb-developers
>> >
>>
>> ------------------------------------------------------------------------------
>> This SF.Net email is sponsored by the Verizon Developer Community
>> Take advantage of Verizon's best-in-class app development support
>> A streamlined, 14 day to market process makes app distribution fast and easy
>> Join now and get one step closer to millions of Verizon customers
>> http://p.sf.net/sfu/verizon-dev2dev
>> _______________________________________________
>> Monetdb-developers mailing list
>> Monetdb-developers at lists.sourceforge.net
>> https://lists.sourceforge.net/lists/listinfo/monetdb-developers
>>
>
> --
> | Dr. Stefan Manegold | mailto:Stefan.Manegold at cwi.nl |
> | CWI,  P.O.Box 94079 | http://www.cwi.nl/~manegold/  |
> | 1090 GB Amsterdam   | Tel.: +31 (20) 592-4212       |
> | The Netherlands     | Fax : +31 (20) 592-4199       |
>




More information about the developers-list mailing list