Comments on: to divide or to multiply

By: Laurent Schneider

Laurent Schneider — Thu, 21 Jun 2007 12:17:47 +0000

> SQL> exec while (:y>1e-125) loop :y:=:y*:z; end loop
>
> PL/SQL procedure successfully completed.
>
> Elapsed: 00:00:00.00

00.00 seems a bit fast, are you sure y was set to 1e125 at the beginning of the test?

By: Laurent Schneider

Laurent Schneider — Tue, 19 Jun 2007 10:41:34 +0000

thanks all for your comments… I was sceptic as I saw this tuning hint. By achieving 40% performance I thought it was worth mentioning. Well, since on APC notebook (is it a notebook with solaris/sparc? ) the division is faster, my test is perhaps is meaningless. Or is it due to TextEdit

It is also dependent from Oracle. Here one more demo with binary_double, which is about 10x faster … the difference is smaller, here on solaris/10.2


SQL> declare
  2    z number;
  3    y number;
  4  begin
  5    z := power(2,103)*1e-31;
  6    for i in 1..10 loop
  7     y := 1e125;
  8     while (y<1e-125) loop y:=y/z; end loop;
  9    end loop;
 10  end;
 11  /
Elapsed: 00:00:02.95
SQL> declare
  2    z number;
  3    y number;
  4  begin
  5    z := power(2,-103)*1e31;
  6    for i in 1..10 loop
  7     y := 1e125;
  8     while (y>1e-125) loop y:=y*z; end loop;
  9    end loop;
 10  end;
 11  /
Elapsed: 00:00:01.54
SQL> declare
  2    z binary_double;
  3    y binary_double;
  4  begin
  5    z := power(2,103)*1e-31;
  6    for i in 1..100 loop
  7     y := 1e125;
  8     while (y>1e-125) loop y:=y/z; end loop;
  9    end loop;
 10  end;
 11  /
Elapsed: 00:00:02.87
SQL> declare
  2    z binary_double;
  3    y binary_double;
  4  begin
  5    z := power(2,-103)*1e31;
  6    for i in 1..100 loop
  7     y := 1e125;
  8     while (y>1e-125) loop y:=y*z; end loop;
  9    end loop;
 10  end;
 11  /
Elapsed: 00:00:02.40

By: Sidhu

Sidhu — Tue, 19 Jun 2007 09:16:11 +0000

BTW if we discuss the reasons for the same…what can be the reasons that multiplication involves lesser processing time than division. One as Laurent said “checking for divide by zero things”. Any other hits…

Sidhu

By: Sidhu

Sidhu — Tue, 19 Jun 2007 09:14:31 +0000

Right Andrew

I too did some RnD on the net. At many places, there are talks about this thing (ie multiplication is faster than division). As you said ultimately it depends upon how the same thing is being done at the resigters’ level.

So can we conclude that its independent of Oracle ? (I mean its entirely dependent on hardware, no way related with Oracle)

Sidhu

By: APC

APC — Tue, 19 Jun 2007 07:51:17 +0000

I was discussing this with some colleagues and the consensus is that it really comes down to chip architecture. Different chips will implement the instructions in different fashions. As it happens I have just read this on the WorseThanFailure site:

“Stephen writes primarily C code …. for tiny devices called “Nodes” that have a whopping 512K of memory, a 2×20 display, and a RISC processor that can’t divide. Really: their compiler treats “%” and “/” as syntactical sugar that gets replaced with a call to a library function which, in turn, uses a series of bitshifts and subtractions to perform division.”

Cheers, APC

By: Eriks

Eriks — Tue, 19 Jun 2007 06:06:59 +0000

That’s a bold statement, knowing that oracle runs not only on windows.

Oracle9i Enterprise Edition Release 9.2.0.7.0 - 64bit Production
PL/SQL Release 9.2.0.7.0 - Production
CORE 9.2.0.7.0 Production
TNS for HPUX: Version 9.2.0.7.0 - Production
NLSRTL Version 9.2.0.7.0 - Production

results are the same for division and multiply.

By: Francois

Francois — Tue, 19 Jun 2007 05:39:57 +0000

Hello,
I think that this can derive from a single question: How basic (processor operations) steps a division takes inside the processor ? Probably much more than for a multiplication.
Francois

By: Sidhu

Sidhu — Mon, 18 Jun 2007 17:27:41 +0000

Laurent

Interesting post

Here are my findings on 10gr2 in my laptop (windows XP)…pretty much in line with you. I ran the same example:

SQL> var z number
SQL> var y number
SQL> exec :z := power(2,102)*2e-31;