#!/bin/sh # Drive the four remaining sweeps sequentially. # # Resource notes (this box: 12 cores, 15 GB RAM, most of it already in use). # An earlier attempt with Pool(11) and an unchunked (n x m) forward pass took # the machine down twice, so: # * CL_NPROC=3 -- at most 3 worker processes at a time # * CL_ZBLOCK -- clcore chunks the (n, m) pre-activation matrix to # 4e6 float64 (32 MB) per block, so peak RSS per # worker is bounded by ~ZBLOCK*8 + m*d*8 regardless # of how large n and m get # * OMP_NUM_THREADS=1 -- no BLAS oversubscription on top of the process pool # The chunked path was verified numerically identical to the unchunked one # (max |diff| 2e-16 on loss, 0 on error rate). # # Each script writes its own results/*.json, so a crash costs at most the one # sweep in flight; finished sweeps are skipped on restart. set -e export OMP_NUM_THREADS=1 export CL_NPROC="${CL_NPROC:-3}" export CL_ZBLOCK="${CL_ZBLOCK:-4000000}" cd "$(dirname "$0")" PY=../.venv/bin/python run_if_missing() { out="results/$1" script="$2" if [ -f "$out" ]; then echo "=== skip $script (have $out)" return 0 fi echo "=== $script $(date +%T)" $PY -u "$script" echo "=== free after $script:" free -g | sed -n 2p } run_if_missing exp2_mechanism.json exp2_mechanism.py run_if_missing exp3_regime.json exp3_regime.py run_if_missing exp4_gengap.json exp4_gengap.py run_if_missing exp5_etaT.json exp5_etaT_needed.py echo "=== ALLDONE $(date +%T)"