Message-ID: <351F9F19.693F@pobox.oleane.com>
Date: Mon, 30 Mar 1998 15:33:13 +0200
From: Francois Charton <deef AT pobox DOT oleane DOT com>
Organization: CCMSA
MIME-Version: 1.0
To: Paul Derbyshire <ao950 AT FreeNet DOT Carleton DOT CA>
CC: djgpp AT delorie DOT com
Subject: Re: Srandom()
References: <6fku51$909 AT freenet-news DOT carleton DOT ca>
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Precedence: bulk

Paul Derbyshire wrote:
> 
> How quickly does srandom() in DJGPP's libc execute? 
>

srand(x) is as fast as can be: it affects the value x to some global 
variable.

For srandom(x), your mileage may vary: there are actually five "flavours" 
of random(x), #defined as TYPE_0 to TYPE_4, the higher the type, the 
randomer the function (the default type used in libc is 3, to change it, 
I think you need to recompile the library).

Type 0 random() is a standard linear congruential generator, srandom(x) 
is the same as srand(). 

For type N, srandom()

1/ initialise N states by calling a rand()-like RNG, N times
2/ avoids "short term effects" by calling random() 10*N times in a row 
(this is the longer part of the program)..

As a result, type N srandom() takes about 10*N times the execution time 
for random() to execute. For the default type 3 random(), srandom() calls 
random() 30 times to initialise.

> Would an added
> srandom() before every random() in a program that uses random()
> extensively cause a noticeable slowdown? 

If it is before every single call to random(), the answer is yes, as soon 
as random() represents a noticeable part of the execution time of your 
program. Anyway, this is clearly a bad idea: the randomness of random() 
is no more used. (if you did that with rand(), calling srand(time()) 
before every occurence of rand(), this would amount to using time() as a 
RNG)

> Reason I ask is because something
> like that might be necessary to make a threadsafe random number generator
> class so that each random number generator instance can be seeded
> independently and give consistent results.
> 

If you want "replicability" of your random series is a multitasking 
environment, each program has to have its own random series. If each 
instance of random() uses long rnadom series (ie several hundred), the 
srandom() overhead will not be too high.

However, if you use such a device, be careful that no program uses random 
numbers from two different series: rand() or random() can guarantee that 
all numbers in a series have certain random properties. It is not clear, 
though, that the numbers from two "parallel" series will be uncorrelated 
(this is wrong in some cases for Linear congruential generators at 
least).

Regards
Francois