Expression timing; a black art?
- To: mathgroup at smc.vnet.net
- Subject: [mg63234] Expression timing; a black art?
- From: AES <siegman at stanford.edu>
- Date: Tue, 20 Dec 2005 04:19:30 -0500 (EST)
- Organization: Stanford University
- Sender: owner-wri-mathgroup at wolfram.com
OK, to follow up on my own recent post, I did some timing tests for the
function I asked about earlier, as a function or as a Module[] with a
repeated square root pulled out. To say the results are puzzling (to
me, anyway) is putting it mildly.
Approach: Create a notebook with three sections, each of the form:
x; Remove["Global`*"];
fn := (function as below)
a=0.12; xmax=3.5; dx=0.1;
tn=Timing[Table[{x, fn[a,x]//N}, {x,0,xmax,dx}]];
where the three functions "fn" are:
f1[a_,x_] := Sum[Exp[-(Pi a n)^2-
((x-n Sqrt[1-(Pi a^2)^2])/a)^2],
{n,-Infinity,Infinity}];
f2[a_,x_] := Module[{b},
b=Sqrt[1-(Pi a^2)^2];
Sum[Exp[-(Pi a n)^2-((x-n b)/a)^2],
{n,-Infinity,Infinity}]];
f3[a_,x_] := Module[{b,c1,c2},
b=Sqrt[1-(Pi a^2)^2];
c1=(Pi a)^2;
c2=1/a^2;
Sum[Exp[-c1 n^2-c2(x-n b)^2],
{n,-Infinity,Infinity}]];
Summary of results to date:
1) Open Mathematica, run notebook first time with functions in f1, f2,
f3 order. Timings are (in round numbers, +/-10%) t1 = 30 sec, t2 = 5
sec, t3 = 20 sec.
2) Re-run same notebook from top: Timings are now 5 sec, 5 sec, 5 sec.
Clearly Mathematica is remembering *something* from the first run, despite the
Remove[Global] in each section . . . ?
3) Quit Mathematica, re-Open, reorder sections in f2, f1, f3 order. Timings on
first run are now t2 = 30 sec, t1 = 5 sec, t3 = 20 sec; timings on
second run are again 5, 5, 5 sec.
4) Quit Mathematica, re-Open, reorder sections in f3, f2, f1 order. Timings on
first run are 30, 20, 5 sec.
Conclusion #1: Running *either* f1 or f2 once leaves something (?) in
the kernel that greatly speeds up the f2 or f1, and gives a little help
to f3. Running f3 first gives a little help to f2 (30 down to 20), and
probably also to f1 (didn't try), but doesn't push it all the way down
to 5.
Conclusion #2: Using modular form with Sqrt[] pulled out doesn't help
at all.
Conclusion #3: If a naive user like me had only done the very first
test above, I'd have been left believing that pulling the Sqrt[] out
*did* help.
Conclusion #4: Trying to understand Mathematica timing is a very black art.
Hypothesis: Running any of these functions on a *random* set of values
the first time, then another random set the second time, will *not*
speed up the second run for either the same fn or any of the other ones.
Anyone want to predict if this is so?
- Follow-Ups:
- Re: Expression timing; a black art?
- From: "Carl K. Woll" <carlw@wolfram.com>
- Re: Expression timing; a black art?
- From: ggroup@sarj.ca
- Re: Expression timing; a black art?