Re: [Gretl-users] Drastic discrepancy in performance

On Sat, 26 Apr 2014, Andreï | Андрей Викторович wrote: I noticed that an attempt to address the cycle variable heavily depends on > the use of dollar sign. Consider the following file [...] > [I'm pushing the example down to insert my comments here.] The issue raised here is the difference in execution speed between variants of a hansl script in which the value of a loop index (say, j) is accessed (variant 1) as simply "j" or (variant 2) in the dollar-form, "$j". Andrei remarks that it "seems more correct" to use "$j", yet it turns out that this slows things down substantially. Short answer: Never use the dollar-form for a loop index unless you really need string substitution. It is always "more correct" to use the plain variable name if possible. Suppose the current value of the index j is 5. Then if you put plain "j" into a command or function call you get the numerical value 5 when the line is evaluated, while if you use "$j" the string "5" is substituted for "$j" in the line in question before it is evaluated (a "macro", so to speak). Consider one of your calls: y = cum(mnormal(T, 1))+5*j What you want here is just the numerical value of j; it would be a roundabout procedure to substitute the string representation of j's value into the line of hansl first, and then evaluate the line. And here's why it slows things down substantially. In the context of loops we attempt to "compile" (in a limited sense) assignments so they'll run faster. But if we find that a given line contains string substitution we abandon the attempt, since in principle the final form of the line could change in an arbitrary, unknown way from one loop iteration to another. Briefly, here's a case where you might want to use $j: series foo_$j = <some expression that depends on j> This gives you a way of defining an arbitrary number of series foo_1, foo_2, and so on, in a loop, something which could not be done conveniently in any other way at present (although we're working on a better mechanism). Allin Cottrell <hansl> #Code for DFtaumu.inp > # Dickey---Fuller's tau_mu distribution > set stopwatch > nulldata 10000 > scalar T = 100 > scalar tranches=10 > scalar iters=10000 > matrix y = zeros(T,1) > matrix DFtaumu = zeros(iters,1) > matrix RES > matrix VARB > > loop j=1..tranches --quiet > loop i=1..iters --quiet > y = cum(mnormal(T, 1))+5*j # <--- That is the one! > DFtaumu[i] = (mols(y[3:], ones(T-2,1)~y[2:T-1], &RES, > &VARB)[2]-1)/VARB[2,2]^0.5 > endloop > sprintf fn "%02d", $j > mwrite(DFtaumu, &quot;DFTM-(a)fn.mat&quot;) > printf "Tranche %d of %d: %f s elapsed\n", $j, tranches, $stopwatch > endloop > </hansl> > > I am studying the influence of the order of magnitude of the constant on > the DF \tau_\mu distribution. The numbers are relatively small for > demonstrative purposes. When I run it through gretlcli or gretlcli-mpi, > the > average running time is 0.2 s per loop of *j* (one tranche). > gretlcli-mpi DFtaumu.inp > However, it seems more correct to use the *$j* instead of just *j* in the > loop, so I modified the 14th line by adding a $: > y = cum(mnormal(T, 1))+5*$j > It seems more correct, yet the performance fell abruptly, the average > running time increased by 50%. When the sample size is increased to > *nulldata 100000* and *scalar iters=100000*, the average execution time > for > one tranche increases from 2.01 (without $ sign) to 2.85 seconds (with $ > sign). > > Why does such a slight change cause such tremendous fall in performance > with seemingly no effect on the properties of the output (I compared the > distributions of DFtaumu and found no difference). Why? What kind of > further optimisation and improvement should be undertaken in order to > avoid > such pitfalls? _______________________________________________ Gretl-users mailing list Gretl-users(a)lists.wfu.edu http://lists.wfu.edu/mailman/listinfo/gretl-users