AMBER Archive (2003)

Subject: Beyond 1,000,000 atoms

From: Vlad (vvv900_at_anusf.anu.edu.au)
Date: Wed Apr 02 2003 - 16:45:21 CST


Concerning the recent buzz on the RNA-RNA complex including 721918
atoms...

CPU time in minutes for various numbers of processors for a system of
1,051,329 atoms, PME, 100 MD time steps with 8 Å cutoff .

1 321.1
2 169.8
4 97.5
8 63.4
16 45.8
32 39.2

CPU time in minutes for various numbers of processors for a
system of 2,162,187 atoms, PME, 100 MD time steps with 8 Å cutoff .

1 695.4
2 368.2
4 211.4
8 132.8

One can see that Sander7 loses some parallel performance for very large
systems. However, profiling showed that increase in the system size leads to
more efficient communication between modules. The communications are
therefore not the problem.

Going from the relatively large system to the huge ones sampling cost of top
subroutines is changing. In the case of very large systems non-bonded list
(pack_nb_list subroutine) preparation takes more time than non-bonded
energy computation (short_ene subroutine). So, decrease in parallel
performance for very large systems was found to arise from the poor serial
performance of Sander7.

All calculation were done on the Fujitsu PrimePower (64 processors running
at 450MHz, 64GB of shared memory).