Gromacs 4 tip tricks-Speeding up your simulations.pdf

Gromacs 4 tips tricks Or: Speeding up your simulations Erik Lindahl lindahl@cbr.su.se CBR Center for Biomembrane Research Stockholm University, Sweden Topics • Algorithms behind the options • Tips tricks to get faster simulations • Simulation cell shapes efficiency • Holonomic constraints, virtual sites, dt • Interaction efficiency, water models • Parallelization, node allocation tricks Some Gromacs stats... • We’re growing - fast! • Gromacs-2.x was ~1Mb compressed • Gromacs-4.x will be ~8Mb compressed • 53.5Mb source code • 1.5 million lines • Now larger than the Linux kernel! • What’s all the new code doing? Why is Gromacs fast? • Algorithmic optimization • Streamlined code: No virial in nonbonded kernels • “Special-purpose software” • Single precision whenever possible • Tuning to avoid conditionals, such as PBC checks in nonbonded kernels (“if”) • Lots of algorithms to support long timesteps and system simplifications Special-purpose code • Coding optimization a0 a1 a2 a3 + • Optimized 1/sqrt(x) b0 b1 b2 b3 • Used ~150,000,000 times/sec = • Handcoded assembly kernels c0 c1 c2 c3 for x86, x86-64, ia64, Altivec, VMX, BlueGene (Fully supported in GROMACS4) • Single instruction, multiple-data • Heavily optimized water interactions • Try hard to avoid calculating null-interactions Interaction efficiency • Avoid calculating zero interactions • Typical water models (SPC, TIP3p,TIP4p) only have LJ on the oxygen • Tabulated LJ expensive, coulomb cheap • 9 atom-atom in