Hi Jed, Indeed, t2-t1=0 always. If I use MPI_Wtime with a MPI program (that uses MPI_Init), this is not a problem and MPI_Wtime() works fine. So it seems to me that for Intel MKL BLACS we may need a different mechanism for measuring the time. I was looking up Intel Trace Collector timing mechanism, but thought I would ask before it gets too complex. https://software.intel.com/en-us/node/561437 Thanks, Tania ________________________________________ From: Jed Brown <[email protected]> Sent: Tuesday, February 21, 2017 8:22 AM To: Hammond, Jeff R; Banerjee,Tania Cc: [email protected] Subject: Re: [JLSE-Intel-NDA] Timing in ScaLAPACK BLACS framework "Hammond, Jeff R" <[email protected]> writes:
MPI wtime returns a double. You are printing as a float so half the bits are lost. Try %lf.
WUT? %f is for doubles. It actually isn't possible to pass float through varargs, they are implicitly promoted. Tania, have you printed t2-t1? The absolute value is of no significance.