adda
adda copied to clipboard
Published
20 hours ago •
adda-team
adda-team
Optimize MPI operations
Several possible ways to optimize MPI part of the code: 1) use one buffer
for all MPI communications; 2) possibly use MPI_ALLTOALL and 'derived
datatypes' for block_transpose; 3) use 'persistent communication requests'
for repeated communications.
Original issue reported on code.google.com by yurkin on 28 Nov 2008 at 6:49
Original comment by yurkin on 10 Jun 2011 at 2:03
- Added labels: MPI
Another large area for possible optimization is file i/o using standard MPI
functions. This can also be used to address issue 90 and issue 31.
#90, #31
Original comment by yurkin on 9 Nov 2011 at 5:31
Another interesting idea is "distributed arrays" data structures instead the
manual distribution, as is done now. The advantage may be that MPI runtime will
be aware of the exact exchanges to be performed inside a single node - they may
be substantially optimized (sometimes even completely omitted). This may also
address issue 137.
#137
Original comment by yurkin on 23 Nov 2011 at 4:09
Probably relevant are the MPI parallelization efforts with DDSCAT, see Numrich RW, Clune TL, Kuo K-S. A new parallel version of the DDSCAT code for electromagnetic scattering from big targets. In: PIERS 2013 Taipei Proceedings; 2013 Mar. 25-28; Taipei, Taiwan. p. 722–726. http://piers.org/piersproceedings/download.php?file=cGllcnMyMDEzVGFpcGVpfDNBM18wNzIyLnBkZnwxMjEwMjAyMTQzMDQ=
In particular, it contains some scaling graphs. Claims very good parallel efficiency for fixed-size problems.
