[Gretl-users] Re: One large matrix or many small matrices for storing simulation results?

On Fri, 23 Apr 2021, Alecos Papadopoulos wrote: I guess it ultimately depends on the nature of your problem: memory allocation has some computational cost, matrix slicing also has one. You have to find the bast tradeoff between the two. Additionally, you should factor in the cost of managing several matrices; for example, the "$i" construct in loops is rather time-consuming, because string operations are involved. Perhaps you could consider matrix arrays as well. For example, try playing with different values for N and K in the following script: <hansl> set verbose off N = 1000 K = 100 H = 100 # 1: one matrix set stopwatch loop H X = zeros(N, K) loop i = 1 .. K X[,i] = 1 endloop endloop t1 = $stopwatch # 2: several matrices set stopwatch loop H loop i = 1 .. K matrix X$i = zeros(N, 1) X$i = 1 endloop endloop t2 = $stopwatch # 3: matrix array set stopwatch loop H matrices XX = array(K) loop i = 1 .. K XX[i] = zeros(N, 1) XX[i] = 1 endloop endloop t3 = $stopwatch print t1 t2 t3 <hansl> Having said all this: in most cases, if the problem is not absolutely trivial, the main factor in determining the speed of your code is likely to be the computational complexity of the main computation, and the way you store things away would only play a marginal role. In a MCMC algorithm such as Metropolis-Hastings, most of your CPU time will probably be spent generating random numbers and computing the likelihood for the various points you visit during the Markov chain no matter how you organise the data. ------------------------------------------------------- Riccardo (Jack) Lucchetti Dipartimento di Scienze Economiche e Sociali (DiSES) Università Politecnica delle Marche (formerly known as Università di Ancona) r.lucchetti(a)univpm.it http://www2.econ.univpm.it/servizi/hpp/lucchetti -------------------------------------------------------