Code execution time measurements - strange

Accelerated Computing HPC Compilers Legacy PGI Compilers
1

Hello,

I’ve got a simple code like this:

	program test

c        Variables
	  integer, save :: nbl, nblcks, ijk, k, l, n, n2
	  integer, dimension(:), allocatable, save :: ijkpr
	  real(kind(zzz)), dimension(:,:), allocatable, save :: vvect
	  real rrnorm

	  real, dimension(2) :: tarray
	  integer :: total, t1, t2, t3, t4, t5, clock_rate, clock_max
	  real :: cputime, gputime

c        Setting the variables values
	  nblcks = 10000000
	  k=10
	  l = 110
	  rrnorm = 3.5

	  allocate(ijkpr(200))
	  allocate(vvect(1000,1000))
	
	  do n=1,200
	     ijkpr(n) = n
	  enddo

	  do n=1,1000
	    do n2=1,1000
	      vvect(n,n2) = n2+n
	    enddo
	  enddo

c        Loop
	  call system_clock(t1, clock_rate, clock_max)
	  do nbl=1,nblcks
	    do n=k,l
	      vvect(n,1)=vvect(n,1)*rrnorm
	    enddo
	  enddo
	  call system_clock(t2, clock_rate, clock_max)
	  cputime = 1000.0*real(t2-t1)/real(clock_rate)
	  write(*,*) 'CPU = ', cputime, ' ms, vvect(100,1)=', vvect(100,1)

	end

And the time results in consecutive code executions are:
CPU = 1125.0 ms
CPU = 0 ms
CPU = 1110.0 ms
CPU = 0 ms
CPU = 0 ms
…

Why sometimes the same code takes 1125 ms and sometimes 0 ms? It is all running on the same machine, same system, just one execution right after another.
I tried different time measurement functions (etime, cpu_time, system_clock, etc) and there is always a similar result - zeros appear sometime. There is the same situation also with totaly different codes.

Thank for clearing me up this problem.

2

Hi szczelba,

Your results will be dependent upon your system’s clock granularity. Given your results, I’m assuming your running on Windows? Windows clock granularity is fairly course.

  • Mat
3

Following your suggestions, I’ve downloaded and ran a program which should change my current timer resolution to about 1ms (called Set Timer Resolution). But still, the results looks like this:

CPU = 1453 ms
CPU = 1438 ms
CPU = 1438 ms
CPU = 0 ms
CPU = 0 ms
…

I read that the default windows xp timer resolution is about 10-20 ms. It still should find a difference between 1453 ms and 0 ms…

How to understand this?