Differences

This shows you the differences between two versions of the page.

Link to this comparison view

Both sides previous revision Previous revision
Next revision		 Previous revision
doku:time [2015/03/26 16:37] – [Timers and performance debugging] irdoku:time [2015/04/17 11:45] (current) – [Hybrid OMP and MPI] ir
Line 5: Line 5:
 [[http://www.mpi-forum.org/docs/mpi-3.0/mpi30-report.pdf|mpi-3.0 Standard]].). [[http://www.mpi-forum.org/docs/mpi-3.0/mpi30-report.pdf|mpi-3.0 Standard]].).
  
-==== MPI_WTIME() ====+==== MPI internal functions ==== 
 + 
 +=== MPI_WTIME() ===
 <code> <code>
 double MPI_Wtime(void) double MPI_Wtime(void)
Line 13: Line 15:
  
 MPI_WTIME returns a floating-point number of seconds, representing elapsed wallclock MPI_WTIME returns a floating-point number of seconds, representing elapsed wallclock
-time since some time in the past and may be used as follows: +time since some time in the past. 
-<code>{ +Generally, **times can ONLY be compared within the same MPI-process.** The times returned are **local to the node that called them**. Di fferent nodes need not to be synchronized. 
-double starttime, endtime; +(MPI_WTIME_IS_GLOBAL is a Boolean variable that is 1 if synchronized and 0 if not. In Open MPI 
-starttime = MPI_Wtime(); +MPI_WTIME_IS_GLOBAL does not have a valid value since the clocks are not guaranteed to be synchronized.)  
-.... stuff to be timed ... + 
-endtime = MPI_Wtime(); +MPI_WTIME has to be called inside the MPI region 
-printf("That took %f seconds\n",endtime-starttime); +<code> 
-}</code> +MPI_Init(&argc, &argv); 
-**Note:** The times returned are **local to the node that called them**. Di fferent nodes need not to be synchronized. +...  
-(MPI_WTIME_IS_GLOBAL is a Boolean variable that is 1 if synchronized and 0 if not.) Therefore, compare times ONLY within the same MPI-process. +MPI_Finalize( ); 
-==== MPI_WTICK() ====+</code> 
 +otherwise its behavior is undefined. 
 + 
 +=== MPI_WTICK() ===
 <code>double MPI_Wtick(void) <code>double MPI_Wtick(void)
 DOUBLE PRECISION MPI_Wtick() BIND(C) DOUBLE PRECISION MPI_Wtick() BIND(C)
Line 29: Line 34:
 MPI_WTICK returns the resolution of MPI_WTIME in seconds, i.e. the time between successive clock ticks given in seconds.  MPI_WTICK returns the resolution of MPI_WTIME in seconds, i.e. the time between successive clock ticks given in seconds. 
  
-==== omp_get_wtime ====+=== Examples === 
 +In this  {{:doku:mpitime.c|example file}}, t1=MPI_WTIME() and t2=MPI_WTIME() enclose the region wich should be timed. 
 +There are two possibilities. Either a MPI_Barrier can be set before each time call. Then the time measured is the time needed that all MPI processes are finished. Or t1 and t2 are defined without a barrier, yielding the execution times for the individual processes. 
 +==== omp internal calls ==== 
 +=== omp_get_wtime ===
  
 c example: c example:
 <code> <code>
-#include "omp.h" 
-#include <stdio.h> 
-#include <windows.h> 
- 
-int main() { 
     double start = omp_get_wtime( );     double start = omp_get_wtime( );
     Sleep(1000);     Sleep(1000);
     double end = omp_get_wtime( );     double end = omp_get_wtime( );
     double wtick = omp_get_wtick( );     double wtick = omp_get_wtick( );
 +</code>
  
-    printf_s("start = %.16g\nend = %.16g\ndiff = %.16g\n", +==== gettimeofday() ==== 
-             start, end, end - start);+gives the time of your wrist watch.
  
-    printf_s("wtick = %.16g\n1/wtick = %.16g\n", +==== clock ==== 
-             wtick, 1.0 / wtick); +gives the amount of time that the CPU is busy for a particular task from one call to the next. 
-}+ 
 +===== Hybrid OMP and MPI ======  
 +{{ :doku:mpi_omp:hybrid.png?400 |}} 
 +{{ :doku:mpi_omp:hybridbarrier.png?400 |}} 
 +{{:doku:mpi_omp:hybridtime.c|}} 
 +====== Serial codes ======  
 +==== fortran ==== 
 +<code> 
 +SYSTEM_CLOCK 
 +DATE_AND_TIME 
 +CPU_TIME
 </code> </code>
 +
  • doku/time.1427387821.txt.gz
  • Last modified: 2015/03/26 16:37
  • by ir