Jim Rosinski (rosinski@cgd.ucar.edu)
Fri, 17 Sep 1999 15:39:30 -0600 (MDT)
On Fri, 17 Sep 1999, Akshara Kaginalkar wrote:
> Can anyone help in result verfication between sequential and parallel
> runs of atmospheric models on distributed memory parallel machine
> using MPI.
> 1. does the result from sequential code and parallel version of the
> same code match perfectly ?
In most cases, yes. Defining the cpp token "SPMD" results in different
memory allocation schemes for some variables however, as well as slightly
different code sequences in a few locations. This can result in roundoff
differences in the (SPMD vs. serial) simulation on some machines.
> 2. If the deviation exists, what is the metorological acceptance
> criteria for it ?
See "The Accumulation of Rounding Errors and Port Validation for Global
Atmospheric Models", Rosinski and Williamson, SIAM J. Sci. Comput, vol 18,
no. 2, pp 552-564, March, 1997.
> 3. does the result from sequential runs from two different
> architectures e.g cray and DEC alpha match exactly ?
No. Different compilers will always generated different instruction
sequences for a code this large. Internal floating point representation and
arithmetic differences will also cause answers to differ.
Regards,
Jim Rosinski
CCM Core Group
This archive was generated by hypermail 2.0b3 on Thu Sep 23 1999 - 12:11:18 MDT