[Monetdb-developers] [Monetdb-pf-checkins] pathfinder/compiler/semantics normalize.brg, 1.11, 1.11.2.1

Jan Rittinger rittinge at in.tum.de
Mon Dec 4 16:49:27 CET 2006


On 12/01/2006 02:28 AM, Peter Boncz wrote with possible deletions:
> The Moerkotte Queries (e.g. /ufs/boncz/moer) seduced me into implementing
> the "syntactic variant" solution to the problem discussed by Jens in the wiki:
> [...]
> I do *not* propose to propagate this to the HEAD automatically. The decision whether
> this rewrite is worth the while I leave to the Munich folks. It is a thing that helps
> mps (a lot, in fact for programmers that use X[Z] XPath notation rather than xquery
> notation for $x in X where Z return $x)
> 

Stefan's automatic propagation however already did exactly that...

(I just played around with both versions and it seems that the algebra 
is not slower with these modifications. In Guidos q11 the modified 
version is even 500 msec faster. My interpretation is that some more 
rewrites on the algebra side are needed to fill the gap :).)

> Anyway, on our standard benchmarking machine db2 (non-optimized 64-bits 32-bits oids
> gcc compile) we now get the following score on the Moerkotte Queries:
> 
>    MXQ.14 MXQ.12 saxon
> q0   0.67   1.71  1.5
> q1   0.26   0.33  1.6
> q2   0.68   1.23   60
> q3   2.38    DNF  209
> q4   1.83   2.02   61
> q5   2.74    DNF  166
> q6   4.82    DNF  130
> q7   2312    DNF  211
> q8    DNF    DNF  DNF
> q9    DNF    DNF  DNF
> q10   DNF    DNF  DNF
> q11   DNF    DNF  151
> q12  0.14    DNF    ?

I tried to get the numbers out of the algebra branch as well. On my 
Laptop (1700MHz, 1024MB Ram) I got better numbers for the first two 
queries compiled with milprint_summer (I only tried these two).

The algebra version compiles for 7 of the 13 queries (3 fail due to non 
existent reverse sorting in the algebra and the 3 others fail as f 
n:number is not implemented yet). Queries q0, q1, q2, q4, q5, q11, and 
q12 compile and in q0, q1, q5, q11, and q12 all joins are recognized (at 
least as far as I could see). q2 and q4 still each contain an undetected 
join because we lack the detection of conjunctive joins (until now).

In the following the measurements of a single(!) run for each query:

guido q0:
alg:
Comp      613.000 msec
Load       51.719 msec
Query     224.477 msec
Print       1.319 msec

mps:
Trans     242.385 msec
Shred      59.625 msec
Query     385.546 msec
Print       1.128 msec

guido q1:
alg:
Comp       79.000 msec
Load       52.124 msec
Query     120.840 msec
Print      16.844 msec

mps:
Trans      62.809 msec
Shred      51.907 msec
Query     155.992 msec
Print      17.736 msec

guido q2 - alg:
Comp      473.000 msec
DNF (dependent join)

guido q4 - alg:
Comp      522.000 msec
DNF (dependent join)

guido q5 - alg:
Comp      03s 255 msec
Load       50.975 msec
Query    1369.837 msec
Print     158.239 msec

guido q11 - alg:
Comp      02s 010 msec
Load       50.960 msec
Query    1130.238 msec
Print      74.269 msec

guido q12 - alg:
Comp      145.000 msec
Load       52.192 msec
Query      53.572 msec
Print       1.980 msec

-- 
Jan Rittinger
Database Systems
Technische Universität München (Germany)
http://www-db.in.tum.de/~rittinge/




More information about the developers-list mailing list