Tiger Team: tt-namd
Performance improvement and benchmarking of NAMD for single-/multinode cases
NAMD is a molecular dynamics package that offers a wide range of configuration options at compile time. It is used in conjunction with quantum mechanics methods on Justus: "Concerning our research on light-harvesting complexes in photosynthesis, we need to use a QM/MM approach for which we use NAMD software to run molecular dynamics calculations. Since these complexes have very high number of atoms (100-200 thousands) we need to use several computer processors in parallel and a positive about NAMD is that it scales very well on multi-nodes simulations."
NAMD performance was evaluated using an example from the user and the apoa1 benchmark (http://www.ks.uiuc.edu/Research/namd/performance.html). For the single-node case, the pre-built NAMD shared-memory binaries needed 71.7 sec (mean of 5 tries, stddev 0.38) on 16 cores, the ones built for Justus with Intel compiler and MKL 36.0 sec (mean of 5 tries, stddev 0.3). Self-built Intel MPI versions were slower with 67.3 sec (mean of 5 tries, stddev 7.8), all other variants were yet slower (not shown). All variants profited from using hyperthreading (32 virtual cores on 16 real ones), bringing down times to 59.9 sec (pre-built shared memory), 31.4 sec (self-built shared memory) and 64.4 sec (self-built Intel MPI). Multinode jobs were possible but not efficient, their run-times for 64 cores were only slightly shorter (down to 24 sec but with large fluctuations on run-time) than those of runs using hyperthreading on a single node. In summary, we could speed up single-node calculations by a factor of 2.3.
Mitglieder des Tiger-Teams: Institut für Theoretische Physik, Universität Ulm; kiz, Universität Ulm