photonics-dhl Octopus Parallel Performance Optimization Skill
Use When
- Running any Octopus GS/TDDFT calculation on HPC
- System has >10 electrons (smaller systems don't benefit from parallelization)
- Available cores: 64-core node
- Want to minimize walltime per calculation
Decision Tree: Which mpirun -np?
System size (electrons)?
├── < 20 e- (e.g. H, H2, CH4)
│ └── mpirun -np = 1 (多核反而更慢，开销大于并行收益)
│
├── 20-50 e- (e.g. ethanol C2H5OH, N2)
│ └── mpirun -np = 8-16
│
├── 50-200 e- (e.g. small molecules, water dimer)
│ └── mpirun -np = 16-32
│
└── > 200 e- (e.g. (H2O)10, proteins, solids)
└── mpirun -np = 32 (最优性价比，效率~46%)
(if time-critical: mpirun -np = 64, 效率~26%)
Key Findings (from benchmark on 64-core node)
| Molecule | Electrons | NP=1 | NP=64 | Speedup | Optimal NP |
|---|---|---|---|---|---|
| CH4 | 10 | 38s | N/A | — | 1 (并行无收益) |
| Ethanol | 26 | 125s | 11s | 11.4x | 8-16 |
| (H2O)10 | 320 | 2310s | 137s | 16.9x | 32 |
PBS Job Template
#!/bin/bash
#PBS -N octopus_calc
#PBS -q workq
#PBS -l nodes=1:ppn=64
#PBS -l walltime=04:00:00
source /data/apps/intel/2018u3/env.sh
WORKDIR="/data/home/zju321/.openclaw/workspace/projects/Dirac/run/bench"
cd "$WORKDIR"
# 每次运行前清理 restart（防止缓存污染时间）
rm -rf restart
export OMP_NUM_THREADS=1
export OCTOPUS_PAR_STATES=64
export OCTOPUS_PAR_DOMAINS=1
export OCTOPUS_PAR_KPOINTS=1
/data/home/zju321/.local/bin/udocker run \
--workdir=/tmp \
--volume="$WORKDIR:/tmp:ro" \
--env=OMP_NUM_THREADS \
--env=OCTOPUS_PAR_STATES \
--env=OCTOPUS_PAR_DOMAINS \
--env=OCTOPUS_PAR_KPOINTS \
--env=LD_LIBRARY_PATH \
bench_octopus \
mpirun -np {NP} octopus > octopus.stdout 2>&1
PP Mode Input: Critical Species Format
%Species
"C" | species_pseudo | set | standard | lmax | 1 | lloc | 0
"H" | species_pseudo | set | standard | lmax | 1 | lloc | 0
"O" | species_pseudo | set | standard | lmax | 1 | lloc | 0
%
DO NOT use file paths in Species block — species_pseudo | set | standard tells Octopus to auto-find the UPF file.
Quick Benchmark Script (7 configs, ~30 min for ethanol)
# Location: run/bench/run_bench_STRONG_SCALING.sh
# PBS: 04:00:00 walltime
# Configs: NP=1,4,8,16,32,48,64
# Always clean restart between runs:
rm -rf restart
# Parse results:
grep '^[ ]*etot =' octopus.stdout | awk '{print $3}' | tail -1
Common Pitfalls
restart 缓存污染 — 不同并行配置共用 restart/ 目录会导致后续配置收敛极快（虚假结果）。每次运行前
rm -rf restart。stdin 重定向失效 —
mpirun octopus < /tmp/inp在 PBS 环境下失败。用 volume mount:--volume="$WORKDIR:/tmp:ro"+--workdir=/tmp，Octopus 自动找/tmp/inp。H 原子 1电子场景 — 不适合并行测例，NP=1 即最优。
NP=64 vs NP=32 — 对 320e- 体系，NP=64 只比 NP=32 快 12%，但浪费 2x 核数。推荐 NP=32。